Journal of Articial Intelligence Research Submitted published

Towards Flexible Teamwork

Milind Tamb e tambeisiedu

Information Sciences Institute and Computer Science Department

University of Southern California

Admiralty Way

Marina del Rey CA USA

Abstract

Many AI researchers are to day striving to build agent teams for complex dynamic

multiagent domains with intended applications in arenas such as education training

entertainment information integration and collective rob otics Unfortunately uncertain

ties in these complex dynamic domains obstruct coherent teamwork In particular team

memb ers often encounter diering incomplete and p ossibly inconsistent views of their en

vironment Furthermore team memb ers can unexp ectedly fail in fullling resp onsibilities

or discover unexp ected opp ortunities Highly exible co ordination and communication is

key in addressing such uncertainties Simply tting individual agents with precomputed

co ordination plans will not do for their inexibility can cause severe failures in teamwork

and their domainsp ecicity hinders reusability

Our central hyp othesis is that the key to such exibility and reusability is providing

agents with general mo dels of teamwork Agents exploit such mo dels to autonomously rea

son ab out co ordination and communication providing requisite exibility Furthermore

the mo dels enable reuse across domains b oth saving implementation eort and enforc

ing consistency This article presents one general implemented mo del of teamwork called

STEAM The basic building blo ck of teamwork in STEAM is joint intentions Cohen

Levesque b teamwork in STEAM is based on agents building up a partial hierar

chy of joint intentions this hierarchy is seen to parallel Grosz Krauss partial Shared

Plans Furthermore in STEAM team memb ers monitor the teams and individual

memb ers p erformance reorganizing the team as necessary Finally decisiontheoretic com

munication selectivity in STEAM ensures reduction in communication overheads of team

work with appropriate sensitivity to the environmental conditions This article describ es

STEAMs application in three dierent complex domains and presents detailed empirical

results

Intro duction

teamwork cooperative eort by the members of a team to achieve a common

goal American Heritage Dictionary

Teamwork is b ecoming increasingly critical in many multiagent environments such as

virtual training Tamb e et al Rao et al interactive education for instance in

virtual historical settings Pimentel Teixeira internetbased information integra

tion Williamson Sycara Decker Rob oCup rob otic and synthetic so ccer Kitano

et al interactive entertainment HayesRoth Brownston Gen Reilly

and p otential multirob otic space missions Teamwork in such complex dynamic

domains is more than a simple union of simultaneous co ordinated activity An illustrative

c

AI Access Foundation and Morgan Kaufmann Publishers All rights reserved

Tambe

example provided by Cohen and Levesque b worth rep eating given that the dier

ence b etween simple co ordination and teamwork is often unacknowledged in the literature

fo cuses on the distinction b etween ordinary trac and driving in a convoy Ordinary

trac is simultaneous and co ordinated by trac signs but it is not considered teamwork

Driving in a convoy however is an example of teamwork The dierence in the two situ

ations is that while teamwork do es involve co ordination in addition it at least involves a

common team goal and co op eration among team memb ers

This article fo cuses on the development of a general mo del of teamwork to enable a

team to act coherently overcoming the uncertainties of complex dynamic environments

In particular in these environments team memb ers often encounter diering incomplete

and p ossibly inconsistent views of the world and mental state of other agents To act

coherently team memb ers must exibly communicate to avoid misco ordination Further

more such environments can often cause particular team memb ers to unexp ectedly fail

in fullling resp onsibilities or to discover unexp ected opp ortunities Teams must thus b e

capable of monitoring p erformance and exibly reorganizing and reallo cating resources to

meet any contingencies Unfortunately implemented multiagent systems often fail to pro

vide the necessary exibility in co ordination and communication for coherent teamwork in

such domains Jennings In particular in these systems agents are supplied

only with preplanned domainsp ecic co ordination When faced with the full brunt of un

certainties of complex dynamic domains the inexibili ty of such preplanned co ordination

leads to drastic failures it is simply dicult to anticipate and preplan for all p ossible con

tingencies Furthermore in scaling up to increasingly complex teamwork situations these

co ordination failures continually recur In addition since co ordination plans are domain

sp ecic they cannot b e reused in other domains Instead co ordination has to b e redesigned

for each new domain

The central hyp othesis in this article is that providing agents with a general mo del

of teamwork enables them to address such diculties Such a mo del enables agents to

autonomously reason ab out co ordination and communication providing them the requisite

exibility in teamwork Such general mo dels also allow reuse of teamwork capabilities

across domains Not only do es such reuse save implementation eort but it also ensures

consistency in teamwork across applications Rich Sidner Fortunately recent

theories of teamwork have b egun to provide the required mo dels for exible reasoning ab out

teamwork eg joint intentions Cohen Levesque b Levesque Cohen Nunes

SharedPlan Grosz Grosz Kraus Grosz Sidner and joint

resp onsibility Jennings are some of the prominent ones among these However

most research eorts have failed to exploit such teamwork theories in building practical

applications Jennings

This article presents an implemented general mo del of teamwork called STEAM sim

ply a Shell for TEAMwork At its core STEAM is based on the joint intentions theory

Levesque et al Cohen Levesque b a but it also parallels and in some

cases b orrows from the SharedPlans theory Grosz Grosz Kraus Grosz

Sidner Thus while STEAM uses joint intentions as the basic building blo ck of team

work as in the SharedPlan theory team memb ers build up a complex hierarchical structure

of joint intentions individual intentions and b eliefs ab out others intentions In STEAM

STEAM co de with do cumentationtraces is available as an online App endix

Towards Flexible Teamwork

communication is driven by commitments emb o died in the joint intentions theory team

memb ers may communicate to attain mutual b elief while building and disbanding joint

intentions Thus joint intentions provide STEAM a principled framework for reasoning

ab out communication providing signicant exibility STEAM also facilitates monitoring

of team p erformance by exploiting explicit representation of team goals and plans If in

dividuals resp onsible for particular subtasks fail in fullling their resp onsibilitie s or if new

tasks are discovered without an appropriate assignment of team memb ers to fulll them

team reorganization can o ccur Such reorganization as well as recovery from failures in

general is also driven by the teams joint intentions

STEAMs op erationalization in complex realworld domains describ ed in the next sec

tion has b een key in its development to address imp ortant teamwork issues discussed ab ove

It has also led STEAM to address some practical issues not addressed in teamwork the

ories One key illustration is in STEAMs detailed attention to communication overheads

and risks which can b e signicant STEAM integrates decision theoretic communication

selectivity agents delib erate up on communication necessities visavis incoherency in

teamwork This decision theoretic framework thus enables improved exibility in commu

nication in resp onse to unexp ected changes in environmental conditions

Op erationalizing general mo dels of teamwork such as STEAM necessitates key mo di

cations in the underlying agent architectures Agent architectures such as Soar Newell

RAP Firby PRS Rao et al BB HayesRoth et al and

IRMA Pollack have so far fo cused on individual agents exible b ehaviors via mech

anisms such as commitments and reactive plans Such architectural mechanisms need to b e

enhanced for exible teamwork In particular an explicit representation of mutual b eliefs

reactive team plans and team goals is essential Additional typ es of commitments suitable

for a team context may need to b e emb o died in the architectures as well Without such

architectural mo orings agents are unable to exploit general mo dels of teamwork and reason

ab out communication and co ordination This view concurs with Grosz who states

that capabilities for teamwork cannot b e patched on but must b e designed in from the

start

Our op erationalization of STEAM is based on enhancements to the Soar architecture

Newell plus a set of ab out domainindep ende nt Soar rules Three dierent

teams have b een develop ed based on this op erationalization of STEAM These teams have

a complex structure of teamsubteam hierarchies and op erate in complex environments

in fact two of them op erate in a commerciallydevelop ed simulation environment for

training This article presents detailed exp erimental results from these teams illustrating

the b enets of STEAM in their development

STEAM is among just a very few implemented general mo dels of teamwork Other

mo dels include Jennings joint responsibility framework in the GRATE system Jennings

based on Joint Intentions theory and Rich and Sidners COLLAGEN Rich Sid

ner based on the SharedPlans theory that b oth op erate in complex domains While

Section will discuss these in greater detail STEAM signicantly diers from b oth these

frameworks via its fo cus on a dierent and arguably wider set of teamwork capabilities

that arise in domains with teams of more than twothree agents with more complex team

organizational hierarchies and with practical emphasis on communication costs

Tambe

The rest of the article b egins with a concrete motivation for our research via a description

of key teamwork problems in realworld domains Section Section discusses theories of

teamwork and sketches their implications for STEAM Section next describ es STEAM our

implemented mo del of teamwork Section discusses STEAMs selective communication

Section presents a detailed exp erimental evaluation Section discusses related work

Finally Section presents summary and future work

Illustrative Domains and Motivations

This investigation fo cuses on three separate domains Two of the domains are based on a

realworld distributed interactive simulator commercially develop ed for military training

Calder et al The simulator enables via networking of several computers

creation of largescale D synthetic battleelds where humans as well as hundreds or even

thousands of intelligent and semiintelligent agents can coparticipate Tamb e et al

The rst domain Attack Figure involves pilot agents for a company of up to eight

synthetic attack helicopters The company starts at the homebase where the commander

pilot agent rst sends orders and instructions to the company memb ers The company

pro cesses these orders and then b egins ying towards their sp ecied battle position ie

the area from which the company will attack the enemy While enroute to the battle

p osition dep ending on the orders the company memb ers may y together or dynamically

split into predetermined subteams Once the company reaches a holding point it halts

One or two scout helicopters y forward and rst scout the battle p osition Based on

communication from the scouts other company memb ers y forward to the battle p osition

Here individual pilots rep eatedly maskhide their helicopters and unmask to sho ot missiles

at enemy targets Once the attack completes the helicopters regroup and return to their

homebase While enroute to the homebase or initially towards the battlep osition if

any company memb er sp ots enemy vehicles p osing a threat to the company it alerts others

The company then evades and bypasses the enemy vehicles while also protecting itself

using guns When the company returns safely to homebase it rearms and refuels readying

itself for the next mission An overview of the overall research and development eort in

this domain simulation infrastructure milestones and agent b ehaviors is presented in Hill et al

HOLDING RIDGE POINT HOME BASE BATTLE ENEMY

POSITION VEHICLES

Figure Attack domain company ying in subteams

In the second domain Transp ort Figure synthetic transp ort helicopters protected

by escort helicopters y synthetic tro ops to land In a typical mission two or four escort

helicopters and four to twelve transp ort helicopters take o from separate ships at sea to

rendezvous at a linkup p oint The escorts then provide a protective cover to the transp ort

helicopters during the entire ight to and from their presp ecied landing zone where the

Towards Flexible Teamwork

synthetic tro ops dismount This domain may involve teams of up to sixteen synthetic pilot

agents the largest team we have encountered although Figure shows twelve

LANDING ZONE

SEA

ESCORT

LAND ESCORT TRANSPORTS ESCORT TRANSPORTS

ESCORT

Figure Transp ort domain with synthetic escort and transp ort helicopters

Our third domain is Rob oCup synthetic so ccer Kitano et al Rob oCup is

an international so ccer tournament for rob ots and synthetic agents aimed at promoting

research in multiagent systems In the synthetic agent track over teams will participate

in the rst Rob oCup tournament at IJCAI in Japan The snapshot in Figure shows

two comp eting teams CMUnited Stone Veloso versus our ISI team

The Attack domain is illustrative of the teamwork challenges In our initial preSTEAM

implementation the helicopter pilot agents were develop ed in the Soar integrated agent

architecture Newell Rosenblo om et al Each pilot agent was based on a

separate copy of Soar For each such pilot an op erator hierarchy was dened Figure

shows a p ortion of this hierarchy Tamb e Schwamb Rosenblo om Op erators

are very similar to reactive plans commonly used in other agent architectures such as

the architectures describ ed in Section Each op erator consists of i precondition rules

for selection ii rules for application a complex op erator subgoals and iii rules for

termination At any one p oint only one path through this hierarchy is active ie it governs

an individuals b ehavior For teamwork among individuals domainsp ecic co ordination

plans were added as commonly done in other such eorts in this typ e of domain Ra jput

Karr Tidhar Selvestrel Heinze Laird Jones Nielsen Coradeschi

including our own Tamb e et al For instance after scouting the battle

p osition a scout executes a plan to inform those waiting b ehind that the battle p osition is

scouted not shown in Figure

Initially with twothree pilot agents and few enemy vehicles limited b ehaviors and con

trolled agent interaction carefully preplanned co ordination was adequate to demonstrate

desired b ehaviors However as the numb ers of agents and vehicles increased their b ehav

iors were enriched and domain exp erts human pilots b egan to sp ecify complex missions

signicant numb ers of unanticipated agent interactions surfaced Faced with the full brunt

Since March a team of graduate students at the Information Sciences Institute ISI has joined

in in further research and maintenance of the ISI team While the author continues to b e resp onsible

for the teamwork in the player agents others have made signicant contributions to individu al agent

b ehaviors

Tambe

Figure The Rob o cup synthetic so ccer domain

EXECUTE−MISSION

Fly−flight−plan Engage Prepare−to return−to−base

Initialize return Mask Unmask Employ to Fly Select Select weapons hover cntrl point route control route point

Dip Initialize Maintain Select− Popup Employ−missile High Low Contour hover masked Mask Goto level level NOE position new−mask ......

location

Figure Attack domain Portion of a pilot agents op erator hierarchy

of the uncertainties in this complex dynamic environment the carefully handco ded pre

planned co ordination led to a variety of teamwork failures in the various demonstrations

and exercises in Figure lists a small sample of the teamwork failures roughly in

the order they were encountered

One approach to address these failures is a further addition of domainsp ecic co ordi

nation plans and indeed this was the rst approach we attempted However there are

several diculties First there is no overarching framework that would enable anticipation

of teamwork failures the teamwork failures just app ear to arise unexp ectedly As a result

Towards Flexible Teamwork

Up on abnormally terminating engagement with the enemy the company commander returned

to home base alone abandoning memb ers of its own company at the battle p osition

Up on reaching the holding area the company waited while a single scout started ying

forward Unfortunately the scout unexp ectedly crashed into a hillside now the rest of the

company just waited indenitely for the scouts scouting message

One pilot agent unexp ectedly pro cessed its initial orders b efore others It then ew towards

the battle p osition while its teammates were left b ehind at the home base

Only a scout made it to the holding area all other helicopters crashed or got shot down

but the scout scouted the battle p osition anyhow and waited indenitely for its nonexistent

company to move forward

When the initial orders unexp ectedly failed to allo cate the scouting role to team memb ers

the company memb ers waited indenitely when they reached the holding p oint

Instructions sent by the commander pilot agent to some company memb ers were lost b ecause

the commander unexp ectedly sent them while the memb ers were busy with other tasks Hence

these memb ers were unable to select appropriate actions

While evading an enemy vehicle encountered enroute one helicopter pilot agent unexp ectedly

destroyed the vehicle via gunre However this pilot agent did not inform others and thus

an unnecessary circuitous bypass route was planned

In an extreme case when all company memb ers ran out of ammunition the company failed

to infer that their mission could not continue

Two separate companies of helicopters were accidentally allowed to use the same radio chan

nels leading to interference and loss of an initial message from one of the company commanders

its company hung indenitely

Figure Some illustrative examples of breakdown in teamwork

co ordination plans have to b e added on a casebycase basis a dicult pro cess since fail

ures have to b e rst encountered in actual runs Furthermore as the system continues to

scale up to increasingly complex teamwork scenarios such failures continually recur Thus

a large numb er of sp ecial case co ordination plans are p otentially necessary Finally it is

dicult to reuse such plans in other domains

Given these diculties we have pursued an alternative approach provide agents

with a general mo del of teamwork The agents can then themselves reason ab out their co

ordinationcommunication resp onsibilitie s as well as anticipate and avoid or recover from

teamwork failures Such an approach also requires an explicit representation of agents team

goals and team plans for that is the very basis for reasoning ab out teamwork Unfortu

nately the agents op erator hierarchy shown in Figure represents its own activities Thus

although the agent is provided information ab out its teammates their participation in par

ticular activities is not explicit but rather implicit in the co ordination plans As a result

the agent remains ignorant as to which op erators truly involve teamwork and the teammates

involved in them For instance executemission and engage are in reality team activities

involving the entire company while mask and unmask involve no teamwork Furthermore

in some team tasks only subteams are involved adding to the diculty of relying on implicit

Tambe

representations since the teammates involved in team tasks vary Even more problematic

for implicit representation are team tasks where the team memb ers p erform nonidentical

activities For instance consider team tasks such as travel ling overwatch where one sub

team travels while the other overwatches or wait while battle position scouted where scouts

scout the battle p osition while the nonscouts wait In such tasks no single agent p erforms

the team activity and yet it is imp ortant to represent and reason ab out the combined ac

tivity that results This diculty in representation is not sp ecic to the Soar architecture

but the entire family of architectures mentioned in Section

More imp ortantly concomitant with the explicit team goals and plans are certain com

mitments and co ordination resp onsibilities towards the team based on the general mo del

of teamwork employed In the absence of b oth the explicit representation of team goals and

plans as well as commitments and resp onsibilities they engender agents are often forced

to rely on the problematic domainsp ecic co ordination plans leading to aforementioned

teamwork failures

Mo dels of Teamwork

Several teamwork theories have b een prop osed in the literature Cohen Levesque b

Grosz Kraus Jennings Kinny et al The theories are not intended

to b e directly implemented say via a theorem prover but to b e used as a sp ecication

for agent design They often prescrib e general rather than domainsp ecic reasoning

pro cesses or heuristics for teamwork Dierent typ es of op erational teamwork mo dels could

p otentially emerge from these theories the space of such mo dels remains to b e fully

explored and understo o d In developing STEAM we have fo cused on the joint intentions

theory Cohen Levesque b Levesque et al Cohen Levesque a given

its detailed formal sp ecication and prescriptive p ower The joint intentions theory is

briey reviewed in Section STEAM ultimately builds on joint intentions in a way that

parallels the SharedPlan theory Grosz Sidner Grosz Grosz Kraus

The SharedPlans theory is very briey reviewed in Section Section sketches the

implications of the theories for STEAM It outlines the rationale for the design decisions

in STEAM in the pro cess it briey compares the capabilities provided by the joint

intentions and SharedPlan theories STEAM is later presented in detail in Sections and

Joint Intentions Theory

The joint intentions framework Cohen Levesque b a Levesque et al

fo cuses on a teams joint mental state called a joint intention A team jointly intends a

team action if team memb ers are jointly committed to completing that team action while

mutually b elieving that they were doing it A joint commitment in turn is dened as a joint

p ersistent goal JPG The team s JPG to achieve p where p stands for completion of a

team action is denoted JPG p q q is an irrelevance clause as describ ed b elow it

enables a team to drop the JPG should they mutually b elieve q to b e false JPG p q

holds i three conditions are satised

All team memb ers mutually b elieve that p is currently false

Towards Flexible Teamwork

All team memb ers have p as their mutual goal ie they mutually know that they want p to

b e eventually true

All team memb ers mutually b elieve that until p is mutually known to b e achieved unachiev

3

able or irrelevant they mutually b elieve that they each hold p as a weak goal WAG WAG

p q where  is a team memb er in implies that one of the following holds

 b elieves p is currently false and wants it to b e eventually true ie p is a normal

achievement goal or

Having privately discovered p to b e achieved unachievable or irrelevant b ecause q is

false  has committed to having this private b elief b ecome s mutual b elief

JPG provides a basic change in plan expressiveness since it builds on a team task p

Furthermore a JPG guarantees that team memb ers cannot decommit until p is mutually

b elieved to b e achieved unachievable or irrelevant Basically JPG p q requires team

memb ers to each hold p as a weak achievement goal WAG WAG p q where is

a team memb er in requires to adopt p as its goal if it b elieves p to b e false However

should privately b elieve that p has terminated ie p is either achieved unachievable

or irrelevant JPGp q is dissolved but is left with a commitment to have this

b elief b ecome s mutual b elief To establish mutual b elief must typically communicate

with its teammates ab out the status of the team task p

The commitment to attain mutual b elief in the termination of p is a key asp ect of a

JPG This commitment ensures that team memb ers stay up dated ab out the status of team

activities and thus do not unnecessarily face risks or waste their time For instance consider

the rst failure presented in Section where the commander returned to home base alone

abandoning its teammates to face a risky situation Such failures can b e avoided given the

commitments in a JPG In our example the commander would have communicated with

its teammates to establish mutual b elief ab out the termination of the engagement

To enter into a joint commitment JPG in the rst place all team memb ers must

establish appropriate mutual b eliefs and commitments An explicit exchange of request

and conrm sp eech acts is one way that a team can achieve appropriate mutual b eliefs

and commitments Smith Cohen Since this exchange leads to establishment of

a JPG we will refer to it in the following as the establish commitments proto col The key

to this proto col is a p ersistent weak achievement goal PWAG PWAG i p denotes

commitment of a team memb er i to its team task p prior to the teams establishing a

JPG initiates the proto col while its teammates in i n resp ond

 executes a Request p cast as an Attempt  That is s ultimate goal  is

to b oth achieve p and have all  i adopt PWAG i p However  is minimally committed

to where denotes achieving mutual b elief in that  has the PWAG to achieve  With

this Request  adopts the PWAG

Each  i resp onds via conrm or refuse Conrm also an Attempt informs others that

 i has the PWAG to achieve p

WAG was originall y called WG in Levesque et al but later termed WAG in Smith Cohen

The PWAG also includes an irrelevance clause q but we will not include it here to simplify the following

description

Tambe

If i  i conrm JPG p is formed

In establishing a JPG this proto col synchronizes In particular with this proto col

memb ers simultaneously enter into a joint commitment towards a current team activity

p While the JPG is the end pro duct of the establish commitment proto col imp ortant

b ehavioral constraints are enforced during execution via the PWAGs In step the adop

tion of a PWAG implies that if after requesting privately b elieves that p is achieved

unachievable or irrelevant it must inform its teammates Furthermore if b elieves that

the minimal commitment is not achieved it must retry eg if a message did not get

through it must retransmit the message Step similarly constrains team memb ers i to

inform others ab out p and to rebroadcast As step indicates all team memb ers must

consent via conrmation to the establishment of a JPG A JPG is not established if any

one agent refuses Negotiations among team memb ers may ensue in such a case however

that remains an op en issue for future work

Shared Plans Theory

In contrast with joint intentions the concept of SharedPlans SP is not based on a joint

mental attitude Grosz Grosz Kraus Grosz Sidner Instead SP

relies on a novel intentional attitude intending that which is similar to an agents normal

intention to do an action However an individual agents intention that is directed towards

its collab orators actions or towards a groups joint action Intention that is dened via a

set of axioms that guide an individual to take actions including communicative actions

that enable or facilitate its teammates subteam or team to p erform assigned tasks Grosz

Kraus

An SP is either a ful l SharedPlan FSP or a partial SharedPlanPSP We will b egin

with a denition of an FSP and then follow with brief remarks ab out a PSP An FSP to

do represents a situation where every asp ect of a joint activity is fully determined

This includes mutual b elief and agreement in the complete recip e R to do R is

a sp ecication of a set of actions which when executed under sp ecied constraints

i

constitutes p erformance of FSPP GR T T R denotes a group GRs plan P

p

at time T to do action at time T using recip e R Very briey FSPP GR T

p p

T R holds i the following conditions are satised

All memb ers of group GR mutually b elieve that they each intend that the prop osition DoGR

T holds ie that GR do es over time T

All memb ers of GR mutually b elieve that R is the recip e for

For each step in R

i

A subgroup GR GR GR has an FSP for using recip e R GR may only b e

k k i i k

an individual in which case it must have a ful l individual plan an analogue of FSP for

individuals

Other memb ers of GR b elieve that there exists a recip e such that GR can bring ab out

k

and have an FSP for but other memb ers may not know R

i i i

For the sake of brevity a context clause C is deleted from this denition Also in this article we will

not address the contracting case discussed in Grosz Kraus

Towards Flexible Teamwork

Other memb ers of GR intend that GR can bring ab out using some recip e

k i

The SharedPlan theory aspires to describ e the entire web of a teams intentions and

b eliefs when engaged in teamwork In this endeavor an FSP represents a limiting case

usually when engaged in a team activity a team only has a partial SharedPlan PSP The

PSP is a snapshot of the teams mental state in a particular situation in their teamwork

and further communication and planning is often used to fulll the conditions of an FSP

although in dynamic domains the team may never actually form an FSP We fo cus on

three relevant arenas in which partiality may exist in a PSP First the recip e R may b e

only partially sp ecied Certainly in dynamic environments such as the ones of interest

in our work recip es could b e considered to evolve over time as teams reactively decide

the next step based b oth on the context and the current situation For instance in the

Attack domain the helicopter company may react to enemy vehicles seen enroute thus

evolving their recip e According to SP theory team memb er must arrive at mutual b elief

in their next steps For each step in the recip e the relevant subgroup must form a

i i

SharedPlan

Second the teams task allo cation may b e unreconciled eg the agent or group to

p erform particular task may not b e determined In this situation team memb ers intend

that there exist some individual or subgroup to do the task Among actions considered as

a result of the intending that individual s may volunteer to p erform the unreconciled task

or p ersuadeorder others to take over the task

Third individuals or subgroups may not have attained appropriate mutual b eliefs for

forming an FSP leading to communication within the team Communication may also

arise due to agents intention that attitude b oth towards their team goal and towards

teammates activities For instance a team memb ers intention that its team do an action

i

and its b elief that communication of some particular information will enable the team to do

will lead it to communicate that information to the team as long as such communication

i

do es not conict with previous commitments

The Inuence of Teamwork Theories on STEAM

In STEAM joint intentions are used as building blo cks of teamwork Several advantages

accrue due to this use First the commitments in a joint intention b egin to provide a

principled framework for reasoning ab out co ordination and communication in teamwork

Thus this framework b egins to address teamwork failures such as those in Figure Second

the joint commitments in joint intentions provide guidance for monitoring and maintenance

of a team activity ie agents should monitor conditions that cause the team activity to

b e achieved or unachievable or irrelevant and maintain the team activity at least until one

of these conditions arises Third a joint intention leads to an explicit representation of a

team activity and thus facilitates reasoning ab out teamwork In particular as shown later

agents can reason ab out the relationship b etween their team activity and an individuals or

subteams contributions to it

However a single joint intention for a highlevel team goal is insucient to provide

all of these advantages To guarantee coherent teamwork four additional issues must b e

addressed Here the SharedPlans theory helps in analysis of STEAMs approach and in

one case STEAM directly b orrows from SharedPlans A key observation is that analogous

Tambe

to partial SharedPlans STEAM builds up snapshots of the teams mental state but via

joint intentions

The rst issue involves coherence in teamwork team memb ers must pursue a common

solution path in service of their joint intention for the highlevel team goal Indeed as

Jennings observes without such a constraint team memb ers could pursue alternative

solution paths that cancel each other so no progress is made towards The SharedPlan

theory addresses such coherence by stepping b eyond the team memb ers intentions that

towards In addition SharedPlans mandates mutual b elief in a common recip e even if

partial and SharedPlans for individual steps in the common recip e thus generating a

i

recursive hierarchy to ensure coherence

STEAMs approach here parallels that of SharedPlans however it builds on joint inten

tions rather than SharedPlans That is STEAM uses joint intentions as a building blo ck

to hierarchically build up the mental attitude of individual team memb ers and ensure that

team memb ers pursue a common solution path In particular as mentioned earlier in

dynamic domains given reactive plans a recip e R may evolve step by step during execu

tion In STEAM as the recip e evolves if a step requires execution by the entire team

i

STEAM requires that the entire team agree on and form joint intentions to execute it

i

To execute a substep of other joint intentions are formed leading to a hierarchy During

i

the expansion of this hierarchy if a step involves only a subteam then that subteam must

form a joint intention to p erform that step If only an individual is involved in the step it

must form an intention to do that step In general the resulting intention hierarchy evolves

dynamically dep ending on the situations the team encounters

Second Grosz and Kraus discuss the tradeos in the amount of information team

memb ers must maintain ab out teammates activities particularly when a step involves

i

only a subteam or an individual Grosz and Kraus address this tradeo in SharedPlans

as shown in step b in Section requiring that team memb ers know only that a recip e

exists to enable a teammates to p erform its actions but not the details of the recip e

Similarly STEAM requires that in case a step is p erformed by a subteam or just an

i

individual team memb er remaining team memb ers track the subteams joint intention or

the relevant team memb ers intention to p erform the step This intention tracking need

not involve detailed plan recognition eg as in our previous work Tamb e

Instead a team memb er must only b e able to infer that its teammates intend or cannot

or do not intend to execute the step This minimal constraint is necessary b ecause

i

otherwise team memb ers may b e unable to monitor the current status of the team activity

eg that their team activity has fallen apart In addition some information ab out the

dep endency relationship among team memb ers actions is useful in monitoring as discussed

in Section

A third issue is the analogue of the unreconciled case in SharedPlans STEAM forms

a joint intention to replan whenever a teams joint intention for a step is seen to b e

i

unachievable Replanning may lead the team to rst analyze the cause of the initial un

achievability Among other p ossibilities the cause could b e the absence of assignment of

a subtask to a subteam or individual or the failure of the relevant individual or subteam

in p erforming the subtask In such a case each team memb er acts to determine the ap

propriate agent or subteam for p erforming the relevant task As a result an agent can

volunteer itself or suggest to other individuals or subteams to p erform the unassigned task

Towards Flexible Teamwork

Of course the unachievabili ty may b e the result of other causes b esides lack of assignment

replanning must then address this other cause further discussion in Section

A nal issue is generalization of STEAMs communication capabilities via a hybrid ap

proach that combines the prescriptions of the joint intentions approach with some asp ects

of SharedPlans A key observation based on Grosz Kraus is that the communica

tion in joint intentions could p otentially b e arrived at in SharedPlans via axioms dening

intention that For instance consider that a team memb er has obtained private informa

tion ab out the achievement of the teams current team action In joint intentions this

team memb er will seek to attain mutual b elief in the achievement of leading to com

munication In contrast in SharedPlans the team memb ers communication would arise

b ecause i it intends that the team do some action which follows and ii the team

cannot do without all team memb ers b eing aware of achievement of Thus further

rst principles reasoning based on interrelationships among actions is required to derive

relevant communication in SharedPlans but in this instance joint intentions provide for

such communication without the reasoning

In general if the teams termination of one action is essential for the team to p erform

some following action the prescription in joint intentions to attain mutual b elief in

termination of team actions is adequate for relevant communication However in some

cases additional communication based on sp ecic informationdep endency relationships

among actions is also essential For instance the scouts in the Attack domain not only

inform all company memb ers of completion of their scouting activity so the company can

move forward but also the precise co ordinates of enemy lo cation to enable the company

to o ccupy go o d attacking p ositions informationdep endency Such communication could

also b e p otentially derived from the axioms of intention that in SharedPlans but at the

cost of further reasoning

STEAM do es not rely on the rstprinciples reasoning from intention that for its commu

nication relying on the prescriptions of joint intentions instead However STEAM exploits

explicit declaration of informationdep endency relationships among actions for additional

communication Thus when communicating the termination of a team action STEAM

i

checks for any inferred or declared informationdep endency relationships with any following

action The information relevant for is also communicated when attaining mutual

j j

b elief in the termination of As a result based on the sp ecic informationdep endency

i

relationship sp ecied dierent typ es of information are communicated when terminating

Thus the scouts can communicate the lo cation of enemy units when communicating

i

the completion of their scouting given the informationdep endency relationship with the

planning of attacking p ositions If no such relationship is sp ecied or if other relationships

are sp ecied the scouts would communicate dierent information

STEAM thus starts with joint intentions but then builds up hierarchical structures

that parallel the SharedPlans theory particularly partial SharedPlans The result could

b e considered a hybrid mo del of teamwork that b orrows from the strengths of b oth joint

intentions formalization of commitments in building and maintaining joint intentions and

SharedPlans detailed treatment of teams attitudes in complex tasks as well as unrecon

ciled tasks This is of course not the only p ossible hybrid As mentioned earlier further

exploration in the space of teamwork mo dels is clearly essential

Tambe

STEAM

STEAMs basis is in executing hierarchical reactive plans in common with architectures

mentioned in Section The novel asp ects of STEAM relate to its teamwork capabilities

The key novelty in STEAM is team operators reactive team plans When agents develop ed

in STEAM select a team op erator for execution they instantiate a teams joint intentions

Team op erators explicitly express a teams joint activities unlike the regular individual

op erators which express an agents own activities In the hierarchy in Figure op erators

shown in such as Engage are team op erators while others are individual op erators

Team activities such as travel ling overwatch or waiting while battle position scouted are now

easily expressed as team op erators as shown in Figure with activities of individual s or

subteams expressed as children of these op erators Team op erators marked with are

typically executed by subteams in this domain

[]EXECUTE−MISSION [ ] Fly−flight−plan ...... [ Engage ] Wait−while− ...... [ battle−position−scouted ] Fly−control [route ] Mask & Scout wait−for Employ Observe forward] scouting weapons [] [ Travelling * * ...... [ Travelling ] [ Overwatch] ......

High Low Unmask Contour Travelling Mask Employ− level level NOE Travelling missile [[Lead ] Cover ] * * ...... Initialize Maintain Goto−new High Low Contour popup Select− Dip

level level NOE hover position Mask mask−location

Figure Attack domain Portion of mo died op erator hierarchy with team operators

As with individual op erators team op erators also consist of i precondition rules ii

application rules and iii termination rules Whether an op erator is a team op erator or

an individual op erator is dynamically determined In particular when an agent i invokes

an op erator for execution the op erator is annotated with an executing agent which may

b e dynamically determined to b e an individual or subteam or a team If the executing

agent is a particular team or subteam the op erator is determined to b e a team op erator

If the executing agent is the agent i itself then an individual op erator results Thus

precise team executing a team op erator is not compiled in but can b e exibly determined

at execution time Figure thus illustrates the conguration of op erators that is typical in

the Attack domain

Given an arbitrary team op erator OP all team memb ers must simultaneously select OP

to establish a joint intention joint intention for OP will b e denoted as OP In Figure

at the highest level the team forms a joint intention for executemission In service of

this joint intention the team may form a joint intention engage In service of engage

individual team memb ers all select individual op erators to employweap ons thus forming

individual intentions An entire hierarchy of joint and individual intentions is thus formed

when an agent participates in teamwork

Towards Flexible Teamwork

A STEAMbased agent maintains its own private state for the application of its indi

vidual op erators and a team state to apply team op erators A team state is the agents

abstract mo del of the teams mutual b eliefs ab out the world eg in the Transp ort do

main the team state includes the co ordinates of the landing zone The team state is usually

initialized with information ab out the team such as the team memb ers in the team p ossible

subteams available communication channels for the team the predetermined team leader

and so forth STEAM can also maintain subteam states for subteam participation There

is of course no shared memory and thus each team memb er maintains its own copy of the

team state and any subteam states for subteams it participates in To preserve the con

sistency of a subteam state one key restriction is imp osed for mo dications to it only

the team op erators representing that subteams joint intentions can mo dify it Thus the

state corresp onding to a subteam can only b e mo died in the context of a joint intention

OP

Thus at minimum STEAM requires the following mo dications to the architectures

such as Soar RAP PRS and others mentioned in Section to supp ort teamwork i gen

eralization of op erators reactive plans to represent team op erators reactive team plans

ii representation of team andor subteam states and iii restrictions on team state mo d

ications only via appropriate team op erators While these team op erators and team

states are at the foundation of STEAM as a general mo del of teamwork STEAM also

involves agents commitments in teamwork monitoring and replanning capabilities and

more Hardwiring this entire teamwork mo del within the agent architectures could p o

tentially lead to unnecessary rigidity in agent b ehaviors Instead the STEAM mo del is

maintained as a domainindep ende nt op erational mo dule eg in the form of rules to

guide agents b ehaviors in teamwork In the future appropriate generalizations of these

capabilities could b egin to b e integrated in agent architectures

The following subsections now discuss key asp ects of STEAM in detail Section

discusses team op erator execution in STEAM Section describ es STEAMs capabilities

for monitoring and replanning Detailed pseudoco de for executing STEAM app ears in

App endix A

Team Op erator Execution

To execute a team op erator agents must rst establish it as a joint intention Thus when

a memb er selects a team op erator for execution it rst executes the establish commitments

proto col describ ed b elow intro duced in Section

Team leader broadcasts a message to the team to establish PWAG to op erator OP Leader

now establishes PWAG If OP not established within time limit rep eat broadcast



Sub ordinates  i in the team wait until they receive leaders message Then turn by turn

broadcast to establishment of PWAG for OP and establish PWAG

Wait until  i  i establish PWAG for OP establish OP



With this establish commitment proto col agents avoid problems of the typ e where just

one memb er ies o to the battle p osition item Figure In particular a team memb er

cannot b egin executing the mission without rst establishing a joint intention execute

mission During execution of the establish commitment proto col PWAGs address several

Tambe

contingencies if an OP is b elieved achieved unachievable or irrelevant prior to OP

agents inform teammates Other contingencies are also addressed eg even if a sub ordinate

initially disagrees with the leader it will conform to the leaders broadcasted choice of

op erators In general resolving disagreements among team memb ers via negotiation is a

signicant research problem in its own right ChuCarroll Carb erry which is not

addressed in STEAM Instead currently STEAM relies on a team leader to initiate the

request and thus resolve disagreements

After establishing a joint intention OP a team op erator can only b e terminated by

up dating the team state mutual b eliefs This restriction on team op erator termination

avoids critical communication failures of the typ e where the commander returned to home

base alone instead agents must now inform teammates when terminating team op erators

Furthermore with each team op erator multiple termination conditions may b e sp ecied

ie conditions that make the op erator achieved unachievable or irrelevant Now if an

agents private state contains a b elief that matches with a team op erators termination

condition and such a b elief is absent in its team state then it creates a communicative

goal ie a communication op erator This op erator broadcasts the b elief to the team

up dating the team state and then terminating the team op erator

As mentioned earlier during teamwork an agent may b e a participant in several joint

intentions some involving the entire team some only a subteam Thus an agent may b e

participating in a joint intention involving the entire company such as executemission

as well as one involving just a subteam such as maskandobserve When the termina

tion condition of a sp ecic team op erator is satised a STEAMbased agent will aim to

attain mutual b elief in only the relevant subteam or team eg facts relevant to maskand

observe may only b e communicated among

During the broadcast of the communication message STEAM checks for information

dep endency relationships with any following tasks if one exists relevant information is

extracted from the current world state and broadcast as well The informationdep endency

relationship may b e sp ecied individual ly p er sp ecic termination condition For instance

supp ose a company memb er sees some enemy tanks on the route while ying to home

base It recognizes that this fact causes the teams current joint intention yightplan

to b e unachievable If this fact is absent in the team state then a communication op erator

is executed resulting in a message broadcast indicating termination of the yightplan

team op erator In addition STEAM uses the explicitly sp ecied informationdep endency

relationship with a following op erator evade to extract the xy lo cation and direction of the

tank As a result the following communication is generated

terminateJPG yightplan evade tank elaborations right

This message identies the sp eaker and informs team memb ers to terminate y

ightplan in order to evade a tank Thus informs others it do es not evade tanks

on its own The part of s message that follows the key word elaborations is due to the

informationdep endency relationship This information the xy lo cation and direction of

the tank enables team memb ers to evade appropriately Separating out the information

Towards Flexible Teamwork

dep endency comp onent in this fashion provides additional communication exibili ty as

explained earlier in Section

Monitoring and Replanning

One ma jor source of teamwork failures as outlined in Section is agents inability to

monitor team p erformance STEAM facilitates such monitoring by exploiting its explicit

representation of team op erators In particular STEAM allows an explicit sp ecication of

monitoring conditions to determine achievement unachievabili ty or irrelevancy of team op

erators In addition STEAM facilitates explicit sp ecication of the relationship b etween a

team op erator and individuals or subteams contributions to it STEAM uses these sp eci

cations to infer the achievement or unachievabili ty of a team op erator These sp ecications

are based on the notion of a role A role is an abstract sp ecication of the set of activities

an individual or a subteam undertakes in service of the teams overall activity Thus a role

constrains a team memb er i or a subteam to some sub op erators op of the team

i

op erator OP For instance supp ose a subteam is assigned the role of a scout in the

Attack domain This role constrains the subteam to execute the sub op erators to scout

the battle p osition in service of the overall team op erator waitwhilebattlepositionscouted

see Figure

Based on the notion of roles three primitive rolerelationships i ANDcombination

ii ORcombination and iii Roledependency can currently b e sp ecied in STEAM These

primitive rolerelationships called rolemonitoring constraints imply the following

relationships b etween a team op erator OP and its sub op erators

V

n

ANDcombination OP op

 i

i=1

W

n

ORcombination OP op

 i

i=1

Role dependency op op op dep endent on op

i j i j

These primitive rolemonitoring constraints may b e combined to sp ecify more com

plex relationships For instance for three agents i j and k with roles op op and

i j

W V

op op op a combination ANDOR role relationship can b e sp ecied as op

j k k i

STEAMbased agents can now infer that the role nonp erformance of k op makes

k

OP unachievable but the role nonp erformance of just one of i or j is not critical to

OP Similarly for two agents i and j b oth an ORcombination plus roledep endency

W V

may b e sp ecied as op op op op Role monitoring constraints may

i j i j

b e sp ecied in terms of individuals roles or subteams roles

The mechanisms for tracking teammates role p erformance or inferring their role non

p erformance is partly domain dep endent As mentioned in Section in some domains

an agent need not know its teammates detailed plan or track that in detail but may rely

on highlevel observations For instance in the Attack domain if a helicopter is destroyed

team memb ers infer role nonp erformance for the aected team memb er In other cases

such as the Rob oCup So ccer domain no such highlevel indication is available Instead

In the future to enable STEAMbased agents to communicate with nonSTEAMbased agents a generic

communication language may b e necessary While generating natural language is currently outside the

scop e of STEAM STEAM do es not preclude such a p ossibil ity Alternatively an articial communication

language such as Sidner may b e used

Tambe

agents need to obtain role p erformance information via agent tracking plan recognition

Tamb e eg is a player agent in the Rob oCup simulation dashing ahead to

receive a pass Communication may b e another source of information regarding role non

p erformance First as discussed b elow STEAM leads individuals to announce rolechanges

to the team and thus other team memb ers indirectly infer rolep erformance information

Second as discussed in Section STEAM may lead individuals to directly communicate

their role nonp erformance Additionally a few domainindep endent mechanisms for infer

ring role p erformance are provided in STEAM Thus role nonp erformance is inferred if no

individual or subteam is sp ecied for p erformance of a role as in item Figure Also

if all individuals within a subteam are found incapable of p erforming their roles STEAM

infers the entire subteam cannot p erform its role

If based on the rolemonitoring constraints and the role p erformance information ab out

teammates STEAM infers team op erator OP to b e unachievable it invokes repair for

replanning By casting repair as a team op erator agents automatically ensure the entire

teams commitment for their replanning the entire team is aected if OP is unachiev

able Furthermore agents inform teammates not only ab out p ossible repair results but

also repair unachievabili ty or irrelevancy The actions taken in service of repair dep end

on the context If repair was invoked due to OP s domainsp ecic unachievabili ty

conditions domainsp ecic repair is triggered In contrast if repair was invoked due to

rolemonitoring constraint failures STEAM leads each agent to rst analyze the failure

The analysis may reveal a critical role failure a single role failure causing the unachiev

ability of OP which may o ccur in an ANDcombination if any agent or subteam fails

in its role or an ORcombination when all team memb ers are roledep endent on a sin

gle individual or a single subteam For instance when agents are ying in formation via

travelling ORcombination everyone is roledep endent on the lead helicopter Thus

should the lead crash a critical role failure o ccurs

The action taken in cases of a critical role failure is team reconguration to determine

a team memb er or subteam to substitute for the critical role As mentioned earlier this

situation corresp onds to the unreconciled case in SharedPlans discussed in Section

The steps taken in STEAM in this case are as follows

Determine candidates for substitution Each team memb er rst matches it own capabilities

or those of other agents or subteams with the requirements of the critical role Matching

currently relies on domainsp ecic knowledge Of course agents or subteams that are the

cause of the critical role failure cannot b e candidates for substitution

Check for critical conicting commitments Once an agent determines p ossible candidates

including itself it checks for conicts with candidates existing commitments to the team If

these commitments are already critical the candidate is eliminated from consideration For

instance if the candidate is a participant in a team op erator which is an ANDcombination

its resp onsibilities to the team are already critical even if it p ossesses relevant capabilities

it cannot take over the role in question Similarly the candidate is ruled out if all other team

memb ers are roledep endent on it

Announce rolesubstitution to the team Candidates not ruled out in step can substitute

for the role This could mean an individual volunteering itself or a team leader volunteering

its subteam for the critical role Since repair is a team op erator and since rolesubstitution



implies its achievement any rolesubstitution is announced to

Towards Flexible Teamwork

Delete noncritical conicting commitments After assuming the new role in the team activity

the relevant individual or subteam memb ers delete their old roles and old commitments

In the Attack domain team memb ers can follow the ab ove pro cedure when recovering

from critical role failures such as item in Figure There since a scouting subteam is

not sp ecied and the relevant op erator waitwhilebattlepositionscouted involves an AND

combination of the scouting role with the nonscouts a critical role failure o ccurs A

subteam in the rest of the company is lo cated to p ossess the capabilities of scouting The

leader of this subteam determines that it can volunteer its subteam for scouting and an

nounces this change in role to the rest of the team Memb ers of this subteam then delete

conicting commitments waitwhilebattlep ositionscouted is now executed with this

new role assignment Since such new role assignments are conned to the lo cal context of

individual team op erators and since step explicitly checks for critical conicts they do

not lead to any global sideeects

The entire repair pro cedure ab ove can invoked in the context of a subteam rather than

the team In this case repair will b e invoked as a team op erator STEAM follows

an identical repair pro cedure in this case enabling individuals or subsubteams to take

over particular critical roles Furthermore any repair communication here is automatically

restricted within

In case the failure is a pure role dependency failure only a single dep endent agent i

is disabled from role p erformance b ecause op op Here i must lo cate another

i j

agent k such that op op Role dep endency failure could involve a subteam

i k i

instead of an individual and the subteams engage in an identical repair

If failure typ e is al l roles failure no agent p erforms its role this state is irreparable

In this situation or in case no substitution is available for a critical role repair is itself

unachievable Since the repair of OP is itself unachievable a complete failure is assumed

and completefailure is now invoked For instance in the Attack domain complete

failure implies returning to home base By casting completefailure as a team op erator

STEAM ensures that team memb ers will not execute such drastic actions without consulting

teammates If only a subteam or an individual i encounters completefailure they infer

inability to p erform their roles in the team s ongoing activity

STEAM Selective Communication

STEAM agents communicate to establish and terminate team op erators Given the large

numb er of team op erators in a dynamic environment this communication is a very sig

nicant overhead as Section shows empirically or risk eg in hostile environments

Therefore STEAM integrates decisiontheoretic communication selectivity Here STEAM

takes into consideration communication costs and b enets as well as the likelihood that some

relevant information may be already mutual ly believed While this pragmatic approach is

a resp onse to the constraints of realworld domains it is not necessarily a violation of the

prescriptions of the joint intentions framework In particular the joint intentions frame

work do es not mandate communication but rather a commitment to attain mutual b elief

Via its decisiontheoretic communication selectivity STEAM attempts to follow the most

costeective metho d of attaining mutual b elief relevant in joint intentions

Tambe

Figure presents the decision tree for the decision to communicate a fact F indicating

the termination of OP Rewards and costs are measured to the team not an individual

The two p ossible actions are NC not communicate cost or C communicate cost Cc

If the action is NC two outcomes are p ossible With probability F was commonly

known anyway and the team is rewarded B for terminating OP With probability

however F was not known and thus there is misco ordination in terminating OP eg

some agents come to know of F only later Given a p enalty C for misco ordination the

mt

reward reduces to BC If the action is C assuming reliable communication F is known mt

Rewards NC (1− ) (F known) B Cost: 0 (F unknown) B − Cmt Decision node 1 (F known) B C Chance node

Cost: Cc 0 (F unknown) B − Cmt

Figure Decision tree for communication

EUC the exp ected utility of option C is BCc EUNC of option NC is B C

mt

To maximize exp ected utility an agent communicates i EUC EUNC ie i

C C c

mt

Thus for instance in the Attack domain when ying with high visibili ty pilot agents do

not inform others of achievement of wayp oints on their route since is low high likeliho o d

of common knowledge and C is low low p enalty However they inform others ab out

mt

enemy tanks on the route since although is low C is high The communication cost

mt

Cc could vary dep ending on the situation as well and team memb ers may exibly reduce

increase communication if the cost increases decreases during their team activity Inter

estingly if only a single agent is left in a team drops to zero and thus no communication

is necessary

Exp ected utility maximization is also used for selectivity in the establish commitments

proto col If is the probability of lack of joint commitments and C the p enalty for

me

executing OP without joint commitments from the team then an agent communicates

i EUC EUNC ie i

C C c

me

Further Communication Generalization

Further generalization in communication is required to handle uncertainty in the termina

tion criteria for joint intentions For instance a team memb er may b e uncertain that an

enemy tank seen enroute causes yightplan to b e unachievable the tanks threat

may not b e clearcut Yet not communicating could b e highly risky The decision tree for

communication is therefore extended to include the uncertainty of an events threat to

the joint intention Figure Since agents may now erroneously inform teammates to

terminate team op erators a nuisance cost Cn is intro duced

Towards Flexible Teamwork

Rewards (Terminates) B (1− ) (1 − ) (Not Terminate) 0 NC (Terminates) B−Cmt Cost: 0 (1 − ) (Not Terminate) 0 (Terminates) B C 1 (1 − ) Cost: Cc (Not Terminate) −Cn

0 [Irrelevant]

Figure Extended decision tree with

Again an agent communicates i EUC EUNC ie i C Cc Cn

mt

If is ie a team op erator has terminated this equation reduces to C Cc

mt

seen previously If ie high uncertainty ab out termination no communication

results if Cn is high Therefore the decision tree is further extended to include a new

message typ e threat to joint intention where Cn is zero but b enets accrued are

lower B C This threat message maximizes exp ected utility when ie if Cn is

high for communicating termination a team memb er communicates a threat For instance

a threat message is used if an agent fails in its own role which is a threat to the joint

intention However as b efore termination messages are used when where they

maximize exp ected utility

Estimating Parameters

As a rst step STEAM only uses qualitative low high medium parameter values STEAM

estimates likeliho o d of lack of joint commitments via team tracking Tamb e

dynamically inferring a teams mental state from observations of team memb ers actions

Fortunately rather than tracking each teammate separately an agent i can rely on its own

team op erator execution for team tracking In particular supp ose i has selected a team

op erator OP for execution and it needs to estimate for op erator OP and its team

Now if i selected OP at random from a choice of equally preferable candidates then its

teammates may dier in this selection Thus there is clearly a low likeliho o d of a joint

commitment i estimates to b e high However if OP is the only choice available then

dep ends on the preceding OP that i executed with the team may b e just a

singleton ie OP may b e an individual op erator that i executed alone There are three

cases to consider First if is subteam of or all memb ers of were

jointly executing OP Furthermore OP could only b e terminated via mutual b elief

among Thus is likely to b e jointly committed to executing the only next choice OP

is estimated low Second if some memb ers in were not jointly participating in

team op erator execution earlier hence is estimated high Third if no op erator precedes

OP eg OP is rst in a subgoal then is estimated low

While agents usually infer matching estimates of sometimes estimates do mismatch

Therefore STEAM integrates some error recovery routines For instance if an agent i

Tambe

estimates to b e low when others estimate it high i starts executing the team op erator

and only later receives messages for establishing joint commitments i recovers by stopping

current activities and reestablishing commitments In contrast if i misestimates to b e

high it unnecessarily waits for messages for establishing commitments STEAM infers such

a misestimation via reception of unexp ected messages it then conducts a lo okahead search

to catch up with teammates

To estimate the probability that a fact is not common knowledge STEAM assumes

identical sensory capabilities for team memb ers eg if some fact is visible to an agent

then it is also visible to all colo cated teammates However at present domain knowledge

is also required to mo del information media such as radio channels in estimating

the probability of an events threat to a joint intention is estimated if a fact matches

sp ecied termination conditions Otherwise role monitoring constraints are used eg in

an ORcombination is inversely prop ortional to the numb er of team memb ers The cost

parameters C C and Cc are assumed to b e domain knowledge

mt me

Evaluation

STEAM is currently implemented within Soar via conventions for enco ding op erators and

states plus a set of rules Essentially these rules help enco de the algorithm in Ap

p endix A some sample rules are presented in App endix B STEAM has b een applied in the

three domains mentioned earlier Attack Transp ort and Rob oCup Table provides some

information ab out the three domains Column lists the three domains Column lists

the maximum numb er of agents p er team in each domain Column shows the p ossible

variations in the sizes of the team Thus in the Attack and Transp ort domains the team

sizes may vary substantially but not so in Rob oCup Column shows the numb er of levels

in the team organization hierarchy eg the teamsubteamindividual hierarchy is a three

level hierarchy Column shows the maximum numb er of subteams active at any one time

Domain Max team Team size Levels in Maximum num

name size varation team hierarchy subteams

Attack

Transp ort

Rob oCup

Table The organizational hierarchy in the three domains

STEAMs application in these three domains provides some evidence of its generality

In particular not only do these domains dier in the team tasks p erformed but as Table

illustrates the domains dier substantially in the team sizes and structure The rest of

this section now uses the three domains in detailed evaluation of STEAM using the criteria

of overall p erformance reusability teamwork exibili ty communication eciency as well

as eort in enco ding and mo difying teamwork capabilities

Towards Flexible Teamwork

Overall Performance

One key evaluation criterion is the overall agentteam p erformance in our three domains

Ultimately STEAMbased agent teams must successfully accomplish their tasks within

their given environments b oth eciently and accurately This is a dicult challenge in

all three domains Certainly the Attack and Transp ort domains involve complex synthetic

military exercises with hundreds of other agents Furthermore in these domains the domain

exp erts exp ert human pilots dene the pilot teams missions tasks rather than the

develop ers STEAMbased pilot teams have so far successfully met the challenges in these

domains they have successfully participated in not one but ab out such synthetic

exercises where the domain exp erts have issued favorable written and verbal p erformance

evaluations

In the Rob oCup domain our player team must comp ete eectively against teams devel

op ed by other researchers worldwide At the time of writing this article our player team

easily wins against the winner of the preRob oCup comp etition However all teams con

tinue to evolve and researchers continue to eld new sophisticated teams One key test for

all the teams in the near future is the Rob oCup tournament at the International Joint

Conference on Articial Intelligence IJCAI Nagoya Japan in August

Reuse of Teamwork Capabilities

STEAMs interdomain and intradomain reusability is approximately measured in Table

Column once again lists the three dierent domains of STEAMs application Column

lists the total numb er of rules p er agent p er domain which include the rules that

enco de the domain knowledge acquired from domain exp erts as well as STEAM rules

illustrating complexity of the agents knowledge base The numb er of STEAM rules used

in these domains is listed in Column Column measures p ercent reuse of STEAM rules

across domains No reuse is shown in STEAMs rst domain Attack There is

reuse in Transp ort ie no new co ordinationcommunication rules were written a ma jor

saving in enco ding this domain Rob oCup in its initial stages has lower reuse Here due

to weakness in spatial reasoning and tracking agents fail to recognize other teams play or

even own teammates failures eg in executing a pass hamp ering the reuse of rules for

rolemonitoring constraints repair and threat detection With improved spatial reasoning

and tracking reuse may improve in the future Column lists the total numb er of team

op erators sp ecied p er domain illustrating signicant intradomain reuse essentially for

each team op erator STEAMs entire teamwork capabilities are brought to b ear

Domain Total rules STEAM rules STEAM reuse Team op erators

Attack rstuse

Transp ort

Rob oCup

Table STEAM reusability data

Tambe

Flexibility in Teamwork

Teamwork exibility is closely related with the measures of overall p erformance and reusabil

ity Since STEAMs entire teamwork capabilities are brought to b ear in executing team

op erators in all of the domains there are signicant improvements in teamwork exibility

For instance in b enchmark runs of Attack almost all of the teamwork failures from our

earlier implementation are avoided Certainly all of the failures in Figure are addressed

Items and are addressed b ecause agents must now attain mutual b elief in the

achievement unachievabili ty or irrelevancy of team op erators Thus in item the

commander now attains mutual b elief that the helicopter company has completed its

engagement with the enemy while in item the irrelevancy of planning a bypass

route is communicated to the company

Items and are addressed b ecause agents now act jointly by rst ensuring the

establishment of joint commitments b efore executing their roles For instance a team

memb er do es not b egin executing the mission as so on as it pro cesses its orders item

rather it acts jointly with the team after the team establishes joint commitments

to execute the mission

Items and are addressed b ecause the team op erator waitwhilebattleposition

scouted is sp ecied to b e an ANDcombination of the role of the scouts and the

nonscouts Thus unachievability of team op erators is detected since either the

scouts or the nonscouts cannot p erform their role or the scoutingrole assignment is

unsp ecied In items and no repairs are p ossible but at the least the company

infers a completefailure and returns to home base instead of waiting indenitely

In item the unassigned role again leads to unachievabili ty but repair is p ossible

b ecause one of the remaining subteams can take over the role of the scout

Item is addressed since the relevant op erator engage is now explicitly dened as a

team op erator with an ORcombination of memb ers roles Thus based on commu

nication from team memb ers team memb ers can infer its unachievabili ty

Item is addressed b ecause in the establishcommitments proto col the leader will

rep eat its message if a resp onse is not heard within time limit However in general

attaining mutual b elief given the p ossibility of uncertain communication channels is

a notoriously dicult challenge Halp ern Moses and this remains an issue

for future work

As a further illustration of teamwork exibility in STEAM we created six variations in

the environmental conditions facing the Attack company of helicopter pilots Each condition

required the pilot team to exibly mo dify its communication to maintain coherence in

teamwork The six variations are

Condition This is the baseline normal condition

Condition Although similar to condition we assume in addition that certain radio

frequencieschannels which were previously separated are now common In partic

ular messages previously assumed to b e privately delivered to only the commander

agent from its sup eriors are now also made available to the other team memb ers

Towards Flexible Teamwork

Condition Although similar to condition the communication cost is raised from

low to medium

Condition Although similar to condition we assume in addition that the he

licopter team has only a medium priority for ensuring simultaneous attack on the

enemy

Condition Here we once again start with the baseline of condition but assume

p o or visibili ty in addition Thus agents may not accurately estimate their distances

Condition In addition to condition here the company has some exibility in

reaching the battle p osition The company is provided with the option of halting at

certain key lo cations rather than continuing to y

The decisiontheoretic framework in STEAM enables agents to exibly resp ond to the

ab ove conditions Figure plots the numb er of messages exchanged among team memb ers

for each of the six conditions The total numb er of messages in three teams balanced

cautious and reckless are compared Balanced agents fully exploit the decision theory

framework and thus illustrate STEAMs exibility Cautious agents always communicate

ignoring the decision theory framework Reckless agents communicate very little only if

high C C Of course truly reckless agents would likely not communicate at all so

mt me

this denition is relaxed here All three teams work with identical cost mo dels Cc C

mt

and C The numb er of agents were xed in this exp eriment to four so all three teams

me

cautious balanced and reckless could b e run as discussed in the next section it is

dicult to run the cautious team with further increase in team size

Fo cusing rst on the balanced team it was able to p erform its mission under all six

conditions by exibly decreasing or increasing the numb er of messages in resp onse The

rst set of conditions conditions through illustrate that the balanced team can reduce

its communication in resp onse to the situation faced eg increase in communication cost

However under conditions and the balanced team can also increase its communication

to address the uncertainties For instance with condition knowledge of p o or visibili ty

automatically leads team memb ers to explicitly communicate achievement of wayp oints

on their route In addition with condition the team has to communicate to establish

commitments when deciding to halt or to y forward

The cautious team was also able to p erform the mission under all six conditions but it

relies on many more messages and remains insensitive to conditions that should result

in fewer messages Indeed its exchange of fold more messages than the balanced

team to p erform an identical task is not only a waste of precious communication resources

but can create risks for the team in hostile environments The next subsection will discuss

the issue of communication eciency in more detail The reckless team do es communicate

fewer messages but it fails to p erform its basic mission Even in the rst normal case

this helicopter company gets stuck on the way to the battle p osition since a message with

medium C but high is not communicated Interestingly the numb er of messages increase

mt

in the reckless team under conditions This is b ecause condition allows the reckless

team to avoid getting stuck b efore reaching the battle p osition In fact this condition was

designed to get the reckless team unstuck Since the reckless team can now p erform more of

Tambe

350 300 250 200 "balanced-u" "cautious-u" 150 "reckless-u" 100

Number of messages 50 0 1 2 3 4 5 6

Types of uncertainties

Figure Change in communication with additional uncertainties

the mission reaching its battle p osition more messages are exchanged Unfortunately

some key messages are still not exchanged leaving team memb ers stranded in the battle

p osition

Communication Eciency

Communication eciency is critical in teamwork particularly with scaleup in team size else

communication overheads can cause signicant degradation in team p erformance Figure

and attempt to measure the communication overhead of teamwork and the usefulness

of STEAMs decisiontheoretic communication selectivity in lowering the overhead partic

ularly for a scaleup in team size Both the gures compare the total numb er of messages

in the three teams intro duced ab ove balanced cautious and reckless with increasing

numb ers of agents p er team In the interest of a fair comparison the total computational

resources available to each team were kept constant a single SUN Ultra While this

limits the maximum team size that could b e run the results shown b elow are suciently

illustrative in terms of scaleup

Figure fo cuses on the Attack domain Decisiontheoretic selectivity enables the bal

anced team to p erform well with few messages this team is regularly elded in synthetic

exercises The cautious team exchanges to fold or more messages than the balanced

team a substantial communication overhead Indeed b eyond six agents the simulation

with cautious team could not b e run in real time Reckless agents in this case do not

exchange any messages at all

Figure fo cuses on the Transp ort domain once again comparing the p erformance of

cautious balanced and reckless teams for increasing numb ers of agents in the team Once

again decisiontheoretic selectivity enables the balanced team to p erform well with few

messages this team is regularly elded in synthetic exercises The cautious team once

again incurs a signicant overhead of to fold or more messages than the balanced

The earlier exp eriments in Section were run with four agents p er team so that the cautious team

could b e run in realtime

Towards Flexible Teamwork

500 450 400 350 300 "balanced" 250 "cautious" "reckless" 200 150 100 Number of messages 50 0 2 3 4 5 6 7 8

Number of agents in team

Figure Attack domain selective communication Reckless team exchanges no messages

and hence that plot overlaps with the xaxis

team Here b eyond seven agents the simulation with the cautious team could not b e

run in real time Interestingly in the test scenario for this exp eriment the reckless team

is able to p erform the mission appropriately even though this team exchanges just

messages far fewer than the balanced team To a certain extent this result illustrates the

p otential for improving the decisiontheoretic selectivity in the balanced team However

when the test scenario for this exp eriment was changed so that the transp orts arrived late

at the rendezvous p oint the balanced team was able to continue to p erform the mission

appropriately However the reckless team now p erformed inappropriately highlighting the

risk in the reckless approach

300

250

200

150

100 "balanced-m" "cautious-m" "reckless-m" Number of messages 50

0 3 4 5 6 7 8

Number of agents in team

Figure Transp ort domain selective communication

Figure illustrates the diering communication patterns in the cautious and balanced

teams for the Attack domain to attempt to understand the dierence in their total com

Tambe

munication Figure a plots the varying degree of collab oration yaxis during dierent

phases xaxis in the Attack domain Degree of collab oration is measured as the p ercentage

of team op erators in a pilots op erator hierarchy which consists of team and individual op

erators A low p ercentage implies low degree of collab oration and vice versa The solid line

plots the overall degree of collab oration in the team taking into account all team op erators

The dashed line indicates the degree of collab oration without counting team op erators exe

cuted by this pilot agents subteam the diering pattern in the two lines is an indication

of the diering degree of subteam activity In particular the two lines sometimes overlap

but separate out at other times indicating the exibility available to the subteam The

overall degree of collab oration is lowest in phases where agents engage the

enemy Figure b plots the percentage of total communication p er phase for cautious

and balanced teams For instance the cautious team exchanges of its total messages in

phase Communication p ercentage is correlated to the degree of collab oration p er phase

for the cautious team co ecient but not for the balanced team co ecient

Essentially unlike the cautious team the balanced team do es not communicate while their

collab oration pro ceeds smo othly

90 25 80 70 20 "cautious" "balanced" 60 15 50 40 10 30 20 5

Degree of Collaboration "team" 10 "team-without-subteam" Percentage of messages 0 0 0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35

Phase Number Phase Number

a Degree of collab oration b Percentage communication

Figure Attack domain pattern of communication

Enco ding and Mo dication Eort

The nal evaluation criteria fo cus on the eort involved in enco ding and mo difying agents

teamwork capabilities comparing the eort in STEAM with alternatives The key alter

native is repro ducing all of STEAMs capabilities via sp ecialcase co ordination plans as in

our initial implementation in the Attack domain We estimate that such an eort would

require signicant additional enco ding eort For example just to repro duce STEAMs

selective communication capabilities our initial implementation could p otentially have re

quired hundreds of sp ecial case op erators Consider our initial implementation in the Attack

domain Here the team op erators in STEAM which would only b e individual op erators

in the initial implementation would each require separate communication op erators two

op erators each to signal commitments request and conrm and one to signal termination

of commitments That is already a total of x Furthermore to repro duce selectivity

additional sp ecial cases would b e necessitated in the extreme case each combination of

values of C and Cc or C and Cc could require a separate sp ecial case op erator

mt me

Towards Flexible Teamwork

x total combinations already more than a hundred Furthermore separate op erators

may b e required dep ending on whether the communication o ccurs with the entire team or

only a subteam Of course it would app ear that all such sp ecial cases could b e economized

in our initial implementation by discovering generalizations but then STEAM enco des

precisely such generalizations to avoid the many sp ecial cases

An additional p oint of evaluation is easy of mo diability of agent team b ehaviors In

our exp erience domain knowledge acquired from exp erts is not static rather it undergo es

a slow evolution In the Attack domain for instance realworld military do ctrine continues

to evolve requiring mo dications in our synthetic pilot team b ehaviors In such situations

STEAM app ears to facilitate such mo dications suggested by domain exp erts at least it

is often not necessary to add new co ordination plans For instance in the Attack domain

domain exp erts earlier suggested a mo dication that the helicopter company should evade

enemy vehicles seen enroute rather than ying over Here adding a new unachievabil

ity condition for the team op erator yightplan was sucient STEAM then ensured

that the pilot agents co ordinated the termination of yightplan even if just one arbi

trary team memb er detected the enemy vehicles Of course the evasion maneuvers b eing

domainsp ecic had to b e added

Related Work

As mentioned earlier most implementations of multiagent collab oration continue to rely on

domainsp ecic co ordination in service of teamwork Jennings More recently

however a few encouraging exceptions have emerged Jennings Rich Sidner

We rst briey review these systems and then contrast them with STEAM

Jenningss implementation of multiagent collab oration in the domain of electric

ity transp ortation management is also based on joint intentions it is likely one of the rst

implementations in a complex domain based on a general mo del of teamwork He presents

a framework called joint responsibility based on a joint commitment to the teams joint

goal and a joint recip e commitment to a common recip e Two distinct typ es of joint

commitments a mo dication to the joint intentions framework are claimed necessary

b ecause dierent actions are invoked when joint commitments are dropp ed However as

a result joint resp onsibility would app ear to b e limited to a twolevel hierarchy of a joint

goal and a joint plan although individuals could execute complex activities in service of

the joint plan The joint resp onsibility framework is implemented in the GRATE system

which app ears to fo cus on a team of three agents In GRATE teamwork pro ceeds with an

organizer agent detecting the need for joint action it is then resp onsible for establishing a

team and ensuring memb ers commitments as required by the joint resp onsibility metho d

While the pro cedure for establishing joint commitments in STEAM is similar to GRATE

including the similarity of the leader in STEAM to the organizer in GRATE

STEAM do es b enet from adopting PWAGs which provides it additional exibility

STEAM is also related to COLLAGEN Rich Sidner a prototyp e to olkit ap

plied to build a collab orative interface agent for applications such as airtravel arrangements

COLLAGENs origins are in the SharedPlans theory Although the COLLAGEN implemen

tation do es not explicitly reason from the intend that attitude in SharedPlans intro duced

in Grosz Kraus it do es incorp orate discourse generation and interpretation algo

Tambe

rithms that originate in such reasoning Lo chbaum Treating the underlying agent as

a blackb ox COLLAGEN facilitates the discourse b etween a human user and the blackb ox

intelligent agent Several COLLAGEN features aid in such interaction such as mainte

nance of a segmented interaction history

STEAM contrasts with COLLAGEN Rich Sidner and GRATE Jennings

in several imp ortant ways First STEAM builds on joint intentions with some

inuence of SharedPlans rather than the SharedPlan approach in COLLAGEN or the

joint resp onsibili ty approach of GRATE Particularly in contrast with joint resp onsibility

STEAM allows teamwork based on deep joint goalplan hierarchies Second STEAM has

the capability for rolemonitoring constraints and role substitution in repairing team activi

ties not relevant in the other two systems Third STEAM has attempted scaleup in team

size Thus STEAM has intro duced techniques b oth to reduce teamwork overheads eg

decisiontheoretic communication selectivity as well as to deal with a hierarchy of teams

and subteams not relevant in smallerscale teams STEAM also illustrates reuse across

domains not seen in the other two systems Finally rather than building a collab oration

layer on top of an existing domainlevel system or blackb ox lo ose coupling STEAM has

prop osed tighter coupling via mo dications to supp ort teamwork in the agent architecture

itself eg with explicit team goals and team states and accompanying commitments The

determining factor here would app ear to b e the tightness of collab oration eg a deeply

nested dynamic joint goal hierarchy should favor a tighter coupling

In our previous work Tamb e b we presented an initial implementation of a team

work mo del also based on joint intentions That work clearly laid the groundwork for

STEAM by dening team op erators and elab orating on their expressiveness However

STEAM was later develop ed b ecause of i several problems in that work in continued de

velopment of teamwork capabilities in the Attack domain ii the presence of new domains

such as Transp ort and iii signicant scaleup in team sizes Since STEAM b oth extends

and substantially revises that earlier work it is b est to treat STEAM as a separate system

rather than an extension of that early work STEAM also provides a conceptual advance

in a clearer analysis and sp ecication of the joint mental attitude it builds up in a team

In particular via an explicit analogy to partial SharedPlans Grosz Kraus this

article has sp elled out the requirement for teams and subteams to build up a hierarchy of

joint intentions b eliefs ab out other team memb ers intentions and joint intentions for the

unreconciled case This analysis also led to a generalization of communication based on

informationdep endency

The following now presents a detailed comparison b etween STEAM and the earlier work

Tamb e b in terms of their capabilities To b egin with STEAM includes an explicit

mechanism to establish joint commitments based on PWAGs which was unaddressed in

previous work so earlier agents would implicitly and hence sometimes incorrectly as

sume the existence of joint commitments Also in earlier work monitoring and repair

was highly sp ecialized In particular the mechanism provided for monitoring was based

on comparing achievement conditions of op erators this mechanism was later discovered

to b e limited to monitoring and repair of just one predetermined sp ecialist role p er team

op erator Furthermore the rolesubstitution was dened via a sp ecial pro cedure executed

separately by individuals In contrast STEAM has signicantly generalized monitoring and

repair via its explicit rolemonitoring constraints that enable monitoring of a much greater

Towards Flexible Teamwork

variety of failures eg the sp ecialist is just one case in all of the varied rolemonitoring

constraint combinations Furthermore STEAM establishes a joint intention to resolve all

failures rather than relying on any sp ecial case pro cedures This is not merely a conceptual

advance in terms of an integrated treatment of repair but has real b ehavioral implications

in providing additional exibility emb o died in the commitments in repair Further

more STEAMs repair generalizes to subteams addresses previous critical commitments

as well as unallo cated tasks In terms of practical concerns our previous work Tamb e

b raised the issue of communication risk in hostile environments but suggested only a

heuristic evaluation of communication costs and b enets a general purp ose mechanism was

lacking STEAM has lled the gap with its decision theoretic framework that now considers

various uncertainties b oth for selective communication as well as enhancements in commu

nication Also unlike STEAM our earlier work did not deal with complex teamsubteam

hierarchies and its mechanisms did not generalize to subteams Finally STEAM is backed

up with detailed exp erimental results ab out both its exibili ty and reuse across domains all

outside the scop e of the previous work

STEAM is also related to co ordination frameworks such as Partial Global PlanningPGP

Durfee Lesser and Generalized Partial Global PlanningGPGP Decker Lesser

Although not driven via theories of collab oration these co ordination frameworks also

strive towards domain indep endence The earlier work on PGP fo cuses on a system of co

op erating agents for consistent interpretation of data from a distributed sensor network

Durfee Lesser Here sub ordinate agents may exchange their individual goals

and plans of action An assigned agent eg a sup ervisor may recognize that individual

plans of dierent agents meld into a partial global plan PGP so called b ecause PGPs

involve more than one agent but not necessarily all agents partially global in service

of a common group goal The PGP is a basis for planning co ordination actions and it

may b e transmitted to sub ordinates for guidance in execution of individual actions PGP

can accommo date dierent typ es of organizations as well GPGP Decker Lesser

provides several indep endent co ordination mo dules any subset of which may b e combined

in resp onse to co ordination needs of a task environment the GPGP approach can duplicate

and extend the PGP algorithm

As a general mo del of teamwork STEAM can provide a principled underlying mo del to

reason ab out at least some of the co ordination sp ecied in PGP eg agents would establish

a joint intention towards the collective goal in a PGP and mo dulate their communication

via decisiontheoretic reasoning That is PGP compiles out some of the underlying rea

soning in STEAM and thus STEAM could provide additional exibility in co ordination

Essentially PGP and GPGP do not separate out co ordination in teamwork from co ordina

tion in general such as via a centralized co ordinator As a result they fail to exploit the

resp onsibilitie s and commitments of teamwork in building up co ordination relationships

Analogously some of the general co ordination in PGP or GPGP is unaccounted for in

STEAM and hence understanding relationships among STEAM and GPGP is an interest

ing area of future work There is a similar relationship b etween STEAM and the COOL

co ordination framework Barbuceanu Fox COOL also fo cuses on general pur

p ose co ordination by relying on notions of obligations among agents However it explicitly

rejects the notion of joint goals and joint commitments It would app ear that individual

Tambe

commitments in COOL would b e inadequate in addressing some teamwork phenomena but

further work is necessary in understanding the relationship among COOL and STEAM

In team tracking Tamb e ie inferring teams joint intentions the expressiveness

of team op erators has b een exploited However issues of establishing joint commitments

communication monitoring and repair are not addressed The formal approach to team

work in Sonenb erg et al transforms team plans into separate roleplans for exe

cution by individual s with rigidly emb edded communications STEAM purp osely avoids

such transformations so agents can exibly reason with i explicit team goalsplans and

ii selective communication seen to b e imp ortant in practice In Gmytrasiewicz Dur

fee Wehe decision theory is applied for message prioritization in co ordination

based on the agents recursive mo deling of each others actions STEAM applies decision

theory for communication selectivity and enhancements but in a very dierent context

practical op erationalization of general domainindep ende nt teamwork mo del based on joint

intentions

Summary and Future Work

Teamwork is b ecoming increasingly critical in a variety of multiagent environments rang

ing from virtual environments for training and education to internetbased information

integration to p otential multirob otic space missions Tamb e et al Rao et al

Pimentel Teixeira Williamson et al Kitano et al HayesRoth et al

Reilly In previous implementations of multiagent systems including our own

teamwork has often b een based on predened domainsp ecic plans for co ordination Un

fortunately these plans are inexible and thus no match for the uncertainties of complex

dynamic environments As a result agents coherent teamwork can quickly dissolve into

misco ordinated misb ehavior Furthermore the co ordination plans cannot b e reused in other

domains Such reuse is imp ortant however b oth to save implementation eort and enforce

consistency across applications

Motivated by the critical need for teamwork exibility and reusability this article has

presented STEAM a general mo del of teamwork While STEAMs development is driven

by practical needs of teamwork applications its core is based on principled theories of team

work STEAM is one of just a few implemented systems that have b egun to bridge the gap

b etween collab oration theories and practice STEAM combines several key novel features

i use of joint intentions as a building blo ck for a teams joint mental attitude Levesque

et al Cohen Levesque b the article illustrates that STEAM builds up

a hierarchical structure of joint intentions and individual intentions analogous to the par

tial SharedPlans Grosz Kraus ii integration of novel techniques for explicit

establishment of joint intentions Smith Cohen iii principled communication

based on commitments in joint intentions iv use of explicit rolemonitoring constraints

as well as repair metho ds based on joint intentions v application of decisiontheoretic

techniques for communication selectivity and enhancements within the context of the joint

intentions framework To avail of the p ower of a mo del such as STEAM a fundamental

change in agent architectures is essential architectures must provide explicit supp ort for

representation of and reasoning with team goals reactive team plans and team states

STEAM has b een applied and evaluated in three complex domains Two of the domains

Towards Flexible Teamwork

Attack and Transp ort are based on a realworld simulation environment for training and

here our pilot agent teams have participated largescale synthetic exercises with hundreds of

other synthetic agents In the third domain Rob oCup our player agent team is now under

development for participation in the forthcoming series of simulated so ccer tournaments

b eginning at IJCAI

Of course STEAM is far from a complete mo del of teamwork and several ma jor is

sues remain op en for future work One key issue is investigating STEAMs interactions

with learning Initial exp eriments with chunking Newell a form of explanation

based learning Mitchell Keller KedarCab elli in STEAM reveal that agents

could automatize routine teamwork activities rather than always reasoning ab out them

Sp ecically from STEAMs domainindep end ent reasoning ab out teamwork agents learn

situationsp ecic co ordination rules For instance when the formation leader crashes an

other agent learns situationsp ecic rules to take over as formation lead and communicate A

wellpracticed team memb er could thus mostly rely on learned rules for routine activities

but fall back on STEAM rules if it encounters any unanticipated situations Additionally

STEAMs knowledgeintensive to learning approach could complement current inductive

learning approaches for multiagent co ordination Sen

Failure detection and recovery is also a key topic for future work particularly in environ

ments with unreliable communication One novel approach exploits agent tracking Tamb e

Rosenblo om Tamb e to infer teammates highlevel goals and intentions for

comparison with own goals and intentions Dierences in goals and intentions may indicate

co ordination failures since teammates often carry out identical or related tasks However

given the overheads of such an approach it has to b e carefully balanced with an agents

other routine activities Initial results of this approach are rep orted in Kaminka Tamb e

Enriching STEAMs communication capabilities in a principled fashion is yet another

key topic for future work Such enriched communication may form the basis of multiagent

collab orative negotiation ChuCarroll Carb erry Currently STEAM relies on the

team or subteam leader when resolving disagreements particularly when deciding the next

action While leadership in teamwork is by itself an interesting phenomena of investigation

enabling agents to negotiate their plans without a leader would also improve STEAMs

exibility We hop e that addressing such issues would ultimately lead STEAM towards

improved exibili ty in teamwork

Acknowledgements

This research was supp orted as part of contract NC from ARPAISO This

article is an extended version of a previous conference pap er Tamb e a I thank Johnny

Chen Jon Gratch Randy Hill and Paul Rosenblo om for their comments and supp ort for

the work rep orted in this article Discussions with Nick Jennings have help ed improve the

quality of the article I also thank ISI team memb ers working on the Rob oCup eort for

their supp ort of the work rep orted in this article Domain exp ertise for this work was

provided by David Sullivan and Greg Jackson of BMH Inc and Wayne Sumner of RDA

Logicon

Tambe

App endix A Detailed STEAM Sp ecication

The pseudoco de describ ed b elow follows the description of STEAM provided in this article

It is based on execution of hierarchical op erators or reactive plans All op erators in the

hierarchy execute in parallel and hence the in parallel construct The comments in the

pseudo co de are enclosed in The terminology is rst describ ed b elow to clarify the

pseudoco de

ExecuteTeamOperator C f ng denotes the execution of a team

op erator by a team given the context of the current intention hierarchy C and

with parameters n

Terms C Cc C are all exactly as in Section

me mt

denotes the team s joint intention to execute

status STATUSOF denotes the status of the joint intention whether

it is mutually b elieved to b e achieved unachievable or irrelevant

satises Achievementconditions f denotes that the fact f satises the achieve

ment conditions of the team op erator similarly with resp ect to unachievabili ty and

irrelevancy conditions

Communicateterminatejpg f denotes communication to the team to termi

nate s joint commitment to due to the fact f

Up datestate teamstate f denotes the up dating of the team state of with

the fact f

Up datestatus denotes the up dating of the team op erator with its current

status of achievement unachievabili ty or irrelevancy

Agent is the individual agent or team executing op erator

actions denote the actions of the op erator

teamtyp e is a test of whether the agent is a team or just one individual

self is a test of whether the agent denotes self

agentstatuschange denotes change in the role p erformance capability of agent

or subteam

ExecuteindividualOperator self C f ng denotes the execution of an

individual op erator by self given the context of the current intention hierarchy C

and with parameters n

For exp ository purp oses Executeteamop erator and Executeindividu alop e rator

are dened as separate pro cedures In reality STEAM do es not dierentiate b etween the

two

Towards Flexible Teamwork

Team Op erator Execution

ExecuteTeamOperator C f ng

f

estimate See Section

if C > Cc execute establish commitments proto col

me

see Section for explanation

establish jointintention



W W

While NOTstatus Achieved status Unachievable status Irrel

  

evant Do

f

W

a if satises Achievementconditions f satises Unachievabilityconditions

W

f satises Irrelevanceconditions f

This is the case where fact f is found to satisfy the termination condition of The

case where f is only a threat to see Section is analogous

f

i estimate  see section

ii if  C > Cc prop oseop erator Communicateterminatejpg f with high

mt

priority

See Section and

iii if no other higher priority op erator in parallel

ExecuteindividualoperatorCommunicateterminatejpg f self C f

g

iv Up datestate teamstate f

v Up datestatus



g

b if agentstatuschange where 

f

i Evaluate rolemonitoring constraints See Section

ii if rolemonitoring constraint failure cf such that satises Unachievabilityconditions

cf then up datestatus



g

c if receive communication of terminatejpg and fact f

f

W

if satises Achievementconditions f satises Unachievabilityconditions

W

f satises Irrelevanceconditions f

f

i Up datestate teamstate f

ii Up datestatus



g

g

d Up datestateteamstate actions

execute domainsp ecic actions to mo dify team state of

Tambe

e if children op erator n of prop osed as candidates

f

i i selectb estf ng

V

ii if teamtyp eAgent i Agent i then in parallel

Executeteamoperator i C fg

V

iii if teamtyp eAgent i Agent i then in parallel

f

A Executeteamoperator i Agent i C fg

B Instantiate rolemonitoring constraints

g

iv if selfAgent i then in parallel

f

A Executeindividualoperator i self C 

B Instantiate rolemonitoring constraints

g

g

g End while statement in

terminate joint intention



if status Unachievable



f

if Repair If is not itself Repair

f

ExecuteteamoperatorRepair C f causeofunachievabilityg

Repair is explained in detail in Section Causeofunachievability passed as a parameter

to Repair may b e rolemonitoring constraint violation as in case b or the domainsp ecic

unachievability conditions

g else f

ExecuteteamoperatorCompleteFailure C f causeofunachievabilityg

If Repair is itself unachievable completefailure results as in Section

g

g

g end pro cedure executeteamop erator

Individual Op erator Execution

ExecuteindividualOperator self C f ng

f

establish as an individual intention

W W

While NOTstatus Achieved status Unachievable status Irrelevant

Do

f

W

a if satises Achievementconditions f satises Unachievabilityconditions

W

f satises Irrelevanceconditions f

f

i Up datestate stateself f

Towards Flexible Teamwork

ii Up datestatus

g

b Up datestatestateself actions

execute domainsp ecic actions to mo dify private state

c if new children op erator f ng of prop osed

f

i i selectb estf ng

ii Executeindividualoperator i self C fg

g

g end while statement in

if status Unachievable

f

if Repair

f

ExecuteindividualoperatorRepair self C f causeofunachievabilityg

Repair is explained in detail in Section Causeofunachievability is only domainsp ecic

unachievability condition This is passed as a parameter to repair

g else f

ExecuteindividualoperatorCompleteFailure self C f causeofunachievabilityg

If Repair is itself unachievable completefailure results as in Section

g

g

g end pro cedure executeindividualop erator

App endix B STEAM Sample Rules

The sample rules describ ed b elow follow the description of STEAM provided in this article

and essentially help enco de the algorithm describ ed in App endix A The rules as with the

algorithm in App endix A are based on execution of hierarchical op erators or reactive plans

While the sample rules b elow are describ ed in simplied ifthen form the actual rules are

enco ded in Soar and are available as an online App endix

SAMPLERULECREATECOMMUNICATIVEGOALONACHIEVED

This rule fo cuses on generating a communicative goal

if an agents private state contains a b elief that satises

the achievement condition of a team op erator OP



See section

IF

agent  is private state contains a fact F

AND

fact F matches an achievement condition AC

of a team op erator OP



AND

fact F is not currently mutually b elieved

AND

a communicative goal for F is not already generated

THEN

Tambe

create p ossible communicative goal CG to communicate fact F to team

to terminate OP



SAMPLERULECREATECOMMUNICATIVEGOALONUNACHIEVABLE

This rule is similar to the one ab ove

IF

agent  is private state contains a fact F

AND

fact F matches an unachievability condition UC

of a team op erator OP



AND

fact F is not currently mutually b elieved

AND

a communicative goal for F is not already generated

THEN

create p ossible communicative goal CG to communicate fact F to team

to terminate OP



SAMPLERULEESTIMATEVALUEFORNONCOMMUNICATION

This rule estimates  C for noncommunication

mt

given a communicative goal using the formula from

Section

IF

CG is a p ossible communicative goal to communicate fact F to team

to terminate OP



AND

C is estimated high

mt

AND

 is estimated low

THEN

Estimated value of noncommunication is medium

SAMPLERULEDECISIONONCOMMUNICATION

This rule makes the communication decision using the formula with  C and Cc

mt

from Section

IF

CG is a p ossible communicative goal to communicate fact F to team

to terminate OP



AND

Estimated value of noncommunication for CG is medium

AND

Estimated value of communication for CG is low

THEN

p ost CG as a communicative goal to communicate fact F to team

to terminate OP



SAMPLERULEMONITORUNACHIEVABILITYANDCOMBINATION

This rule checks for unachievability of rolemonitoring

constraints involving an ANDcombination See section

Towards Flexible Teamwork

IF

A current joint intention OP involves an ANDcombination



AND

 i is a memb er p erforming role to execute subop erator op

AND

no other memb er  j is also p erforming role to execute subop erator op

AND

 i cannot p erform role

THEN

Current joint intention OP is unachievable due to a critical role failure



of  i in p erforming op

References

Barbuceanu M Fox M The architecture of an agent building shell In

Wo oldridge M Muller J Tamb e M Eds Intel ligent Agents Volume II Lec

ture Notes in Articial Intel ligence SpringerVerlag Heidelb erg Germany

Calder R B Smith J E Courtemanche A J Mar J M F Ceranowicz A Z

Mo dsaf b ehavior simulation and control In Proceedings of the Conference on

Computer Generated Forces and Behavioral Representation

ChuCarroll J Carb erry S Conict detection and resolution in collab orative

planning In Wo oldridge M Muller J Tamb e M Eds Intel ligent Agents

Volume II Lecture Notes in Articial Intel ligence SpringerVerlag Heidelb erg

Germany

Cohen P R Levesque H J a Conrmation and joint action In Proceedings of

the International Joint Conference on Articial Intel ligence

Cohen P R Levesque H J b Teamwork Nous

Coradeschi S A decision mechanism for reactive and co ordinated agents Tech

rep Linkoping University Licentiate Thesis

Decker K Lesser V Designing a family of co ordination algorithms In Proceed

ings of the International Conference on MultiAgent Systems

Durfee E Lesser V Partial global planning a co ordination framework for

distributed planning IEEE transactions on Systems Man and Cybernetics

Firby J An investigation into reactive planning in complex domains In Proceedings

of the National Conference on Articial Intel ligence AAAI

Gmytrasiewicz P J Durfee E H Wehe D K A decision theoretic approach

to coordinating multiagent interactions In Proceedings of International Joint Con

ference on Articial Intel ligence

Grosz B Collab orating systems AI magazine

Tambe

Grosz B Kraus S Collab orative plans for complex group actions Articial

Intel ligence

Grosz B J Sidner C L Plans for discourse In Cohen P R Morgan

J Pollack M Eds Intentions in Communication pp MIT Press

Cambridge MA

Halp ern J Y Moses Y Knowledge and common knowledge in a distributed

environment Journal of the ACM

HayesRoth B Brownston L Gen R V Multiagent collab oration in directed

improvisation In Proceedings of the International Conference on MultiAgent Systems

ICMAS

Hill R Chen J Gratch J Rosenblo om P Tamb e M Intelligent agents for

the synthetic battleeld a company of rotary wing aircraft In Proceedings of the

Innovative Applications of Articial Intel ligence

Jennings N Commitments and conventions the foundation of co ordination in

multiagent systems The Know ledge Engineering Review

Jennings N Controlling co op erative problem solving in industrial multiagent

systems using joint intentions Articial Intel ligence

Kaminka G A Tamb e M So cial comparison for failure monitoring and recov

ery in multiagent settings In Proceedings of the National Conference on Articial

Intel ligence p Student abstract

Kinny D Ljungb erg M Rao A Sonenb erg E Tidhard G Werner E

Planned team activity In Castelfranchi C Werner E Eds Articial Social

Systems Lecture notes in AI Springer NY

Kitano H Asada M Kuniyoshi Y No da I Osawa E Rob o cup The rob ot

world cup initiative In Proceedings of IJCAI Workshop on Entertainment and

AIAlife

Kitano H Tamb e M Stone P Veloso M No da I Osawa E Asada M

The rob o cup synthetic agents challenge In Proceedings of the International Joint

Conference on Articial Intel ligence IJCAI

Laird J E Jones R M Nielsen P E Co ordinated b ehavior of computer

generated forces in tacairsoar In Proceedings of the Fourth Conference on Com

puter Generated Forces and Behavioral Representation Orlando Florida Institute

for Simulation and Training University of Central Florida

Levesque H J Cohen P R Nunes J On acting together In Proceedings of

the National Conference on Articial Intel ligence Menlo Park Calif AAAI press

Lo chbaum K E Using col laborative plans to model the intentional structure of

discourse PhD thesis Harvard University

Towards Flexible Teamwork

Mitchell T M Keller R M KedarCab elli S T Explanationbased general

ization A unifying view Machine Learning

Newell A Unied Theories of Cognition Harvard Univ Press Cambridge Mass

Pimentel K Teixeira K Virtual reality Through the new looking glass

WindcrestMcGrawHill Blue Ridge Summit PA

Pollack M The uses of plans Articial Intel ligence

Ra jput S Karr C R Co op erative b ehavior in mo dsaf Tech rep ISTCR

Institute for simulation and training University of Central Florida

Rao A S Lucas A Morley D Selvestrel M Murray G Agentoriented

architecture for aircombat simulation Tech rep Technical Note The Australian

Articial Intelligence Institute

Reilly W S Believable Emotional and Social Agents PhD thesis Scho ol of

Computer Science Carnegie Mellon University

Rich C Sidner C COLLAGEN When agents collab orate with p eople In

Proceedings of the International Conference on Autonomous Agents Agents

Rosenblo om P S Laird J E Newell A McCarl R A preliminary analysis of

the soar architecture as a basis for general intelligence Articial Intel ligence

Sen S Proceedings of the Spring Symposium on Adaptation Coevolution and Learn

ing American Asso ciation for Articial Intelligence Menlo Park CA

Sidner C An articial discourse language for collab orative negotiation In Proceed

ings of the National Conference on Articial Intel ligence AAAI

Smith I Cohen P Towards semantics for an agent communication language

based on sp eech acts In Proceedings of the National Conference on Articial Intel li

gence AAAI

Sonenb erg E Tidhard G Werner E Kinny D Ljungb erg M Rao A

Planned team activity Tech rep Australian AI Institute

Stone P Veloso M Towards collab orative and adversarial learning a case

study in rob otic so ccer In Sen S Ed AAAI Spring Symposium on Adaptation

Coevolution and Learning in multiagent systems

Tamb e M Recursive agent and agentgroup tracking in a realtime dynamic en

vironment In Proceedings of the International Conference on Multiagent systems

ICMAS

Tamb e M Tracking dynamic team activity In Proceedings of the National Con

ference on Articial Intel ligence AAAI

Tambe

Tamb e M a Agent architectures for exible practical teamwork In Proceedings of

the National Conference on Articial Intel ligence AAAI

Tamb e M b Implementing agent teams in dynamic multiagent environments Ap

plied Articial Intel ligence to app ear

Tamb e M Johnson W L Jones R Koss F Laird J E Rosenblo om P S

Schwamb K Intelligent agents for interactive simulation environments AI

Magazine

Tamb e M Rosenblo om P S RESC An approach for realtime dynamic agent

tracking In Proceedings of the International Joint Conference on Articial Intel ligence

IJCAI

Tamb e M Schwamb K Rosenblo om P S Building intelligent pilots for

simulated rotary wing aircraft In Proceedings of the Fifth Conference on Computer

Generated Forces and Behavioral Representation

Tidhar G Selvestrel M Heinze C Mo deling teams and team tactics in

whole air mission mo deling Tech rep Technical Note The Australian Articial

Intelligence Institute

Williamson M Sycara K Decker K Executing decisiontheoretic plans in

multiagent environments In Proceedings of the AAAI Fal l Symposium on Plan Exe

cution Problems and Issues