Multiagent Systems Katia P

Appears in: AI magazine Volume 19, No.2 Intelligent Agents Summer 1998. Articles

Multiagent Systems Katia P. Sycara

Agent-based systems technology has generated lots particularly complex, large, or unpredictable, then the of excitement in recent years because of its promise only way it can reasonably be addressed is to develop as a new paradigm for conceptualizing, designing, a number of functionally specific and (nearly) modular and implementing software systems. This promise is components (agents) that are specialized at solving a particularly attractive for creating software that particular problem aspect. This decomposition allows operates in environments that are distributed and each agent to use the most appropriate paradigm for open, such as the internet Currently, the great solving its particular problem. When interdependent majority of agent-based systems consist of a single problems arise, the agents in the system must agent However as the technology matures and coordinate with one another to ensure that addresses increasingly complex applications, the interdependencies are properly managed. need for systems that consist of multiple agents that Furthermore, real problems involve distributed, communicate in a peer-to-peer fashion is becoming open systems (Hewitt 1986). An open system is one in apparent. Central to the design and effective which the structure of the system itself is capable of operation of such multiagent systems (MASs) are a dynamically changing. The characteristics of such a core set of issues and research questions that have system are that its components are nor known in been studied over the years by the distributed AI advance; can change over time; and can consist of community. In this article, I present some of the highly heterogeneous agents implemented by differ- critical notions in MASs and the research work that ent people, at different times, with different software has addressed them- I organize these notions around tools and techniques. Perhaps the best-known example the concept of problem-solving coherence, which I of a highly open software environment is the internet. believe is one of the most critical overall The internet can be viewed as a large, distributed characteristics that an MAS should exhibit information resource, with nodes on the network designed and implemented by different organizations ost researchers in AI to date have dealt with and individuals. In an open environment, information developing theories, techniques, and sources, communication links, and agents could Msystems to study and understand the appear and disappear unexpectedly. Currently, agents behavior and reasoning on the internet mostly perform information retrieval properties of a single cognitive entity. AI has matured, and filtering. The next generation of agent technology and it endeavors to attack more complex, realistic, and will perform information gathering in context and large-scale problems. Such problems are beyond the sophisticated reasoning in support of user problem- capabilities of an individual agent. The capacity of an solving tasks. These capabilities require that agents be intelligent agent is limited by its knowledge, its able to interoperate and coordinate with each other in computing resources, and its perspective. This bounded peer-to-peer interactions. In addition, these rationality (Simon 1957) is one of the underlying capabilities will allow agents to increase the problem- reasons for creating problem-solving organizations. solving scope of single agents. Such functions will The most powerful tools for handling complexity are require techniques based on negotiation or modularity and abstraction. Multiagent systems cooperation, which lie firmly in the domain of MASs (MASs) offer modularity. If a problem domain is (Jennings, Sycara, and Wooldridge 1998;

The autonomous and interacting with other similar agents O’Hare and Jennings 1996; Bond and Gasser 1988). that manage calendars of different users (Garrido and characteristics It is becoming increasingly clear that to be successful, Sycara 1996; Dent et al. 1992). Such agents also can of MASs are increased research resources and attention should be be customized to reflect the preferences and constraint that given to systems consisting of not one but multiple agents. of their users. Other examples include air-traffic The distributed AI (DAD community that started forming in control (Kinny et al 1992; Cammarata, McArthur, and (1) each the early 1980s and was tiny compared to mainstream, single- Steeb 1983) and multiagent bargaining for buying and agent has agent AI is rapidly increasing. The growth of the MAS field is selling goods on the internet indisputable. Fourth is to provide solutions that efficiently use incomplete Research in MASs is concerned with the study, behavior, information sources that are spatially distributed. information and construction of a collection of possibly preexisting Examples of such domains include sensor networks autonomous agents that interact with each other and their (Corkill and Lesser 1983), seismic monitoring (Mason or capabilities environments. Study of such systems goes beyond the study of and Johnson 1989), and information gathering from for solving individual intelligence to consider, in addition, problem the internet (Sycara et al. 1996). the problem solving that has social components. An MAS can be defined as Fifth is to provide solutions in situations where a loosely coupled network of problem solvers that interact to expertise is distributed. Examples of such problems and, thus, solve problems that are beyond the individual capabilities or include concurrent engineering (Lewis and Sycara has a limited knowledge of each problem solver (Durfee and Lesser 1989). 1993), health care, and manufacturing. These problem solvers, often ailed agent, are autonomous and Sixth is to enhance performance along the viewpoint; (2) can be heterogeneous in nature. dimensions of (1) computational efficiency because there is no The characteristic of MASs are that (1) each agent has concurrency of computation is exploited (as long as incomplete information or capabilities for solving the problem communication is kept minimal, for example, by system global and. thus, has a limited viewpoint; (2) there is no system transmitting high-level information and results rather control; global control; (3) data are decentralized; and (4) computation than low-level data); (2) reliability, that is, graceful is asynchronous. The motivations for the increasing interest in recovery of component failures, because agents with (3) data are MAS research include the ability of MASs to do the redundant capabilities or appropriate interagent decentralized; following: coordination are found dynamically (for example, and (4) First is to solve problems that are too large for a centralized taking up responsibilities of agents that fail); (3) agent to solve because of resource limitations or the sheer risk extensibility because the number and the capabilities of computation of having one centralized system that could be a performance agents working on a problem can be altered; (4) is bottleneck or could fail at critical times. robustness, the system’s ability to tolerate uncertainty Second is to allow for the interconnection and because suitable information is exchanged among asynchronous. interoperation of multiple existing legacy systems. To keep agents; (5) maintainability because a system composed pace with changing business of multiple components-agents is easier to maintain needs, legacy systems must periodically be because of Its modularity; (6) responsiveness because updated. Completely rewriting such software tends to be modularity can handle anomalies locally, not prohibitively expensive and is often simply impossible. propagate them to the whole system; (7) flexibility Therefore, in the short to medium term, the only way that such because agents with different abilities can adaptively legacy systems can remain useful is to incorporate them into a organize to solve the current problem; and wider cooperating agent community in which they can be (8) reuse because functionally specific agents can be exploited by other pieces of software. Incorporating legacy reused in different agent teams to solve different systems into an agent society can be done, for example, by problems. building an agent wrapper around the software to enable it to MASs are now a research reality and are rapidly interoperate with other systems (Genesereth and Ketchpel having a critical presence in many human-computer 1994). environments. My purpose in this article is not to Third is to provide solutions to problems that can naturally be provide a detailed review of the field; I leave this task regarded as a society of autonomous interacting components- to others (see, for example, Huhns and Singh [1997], agents. O’Hare and Jennings [19961, Wooldridge and For example, in meeting scheduling a scheduling agent that Jennings [1995], Chaib-draa et al. [19921, and Bond manages the calendar of its user can be regarded as and

AI MAGAZINE

Gasser [1988] for surveys). Rather than present an in- Ensuring that an MAS exhibits coherent collective depth analysis and critique of the field, I instead briefly behavior while it avoids unpredictable or harmful Sophisticated discuss some key topic and indicate how they are behavior (for example, chaos, oscillation) is indeed a individual interrelated. Where appropriate, references to more major challenge: By its very nature, an MAS lacks detailed treatments are provided. global perspective, global control, or global data. agent Coherence is a global (or regional) property of the reasoning can MAS that could be measured by the efficiency, quality increase MAS Multiagent System and consistency of a global solution (system behavior) Issues and Challenges as well as the ability of the system to degrade coherence Although MASs provide many potential advantages, gracefully in the presence of local failures. Several because each they also present many difficult challenges. Here, I methods for increasing coherence have been studied. present problems inherent in the design and These methods, along with issues of single-agent individual implementation of MASs. The list of challenges structuring in an MAS, cover the topic I want to agent can includes problems first posed in Bond and Gasser survey here. (1988), but I have added some: reason about First, how do we formulate, describe, decompose, Individual Agent Reasoning nonlocal effects and allocate problems and synthesize results among a Sophisticated individual agent reasoning can increase group of intelligent agents? MAS coherence because each individual agent can of local Second, how do we enable agents to communicate reason about nonlocal effects of local actions, form actions, and interact? What communication languages and expectations of the behavior of others, or explain and protocols do we use? How can heterogeneous agents possibly repair conflicts and harmful interactions. form interoperate? What and when can they communicate? Numerous works in AI research try to formalize a expectations of How can we find useful agents in an open logical axiomatization for rational agents (see the behavior of environment? Wooldridge and Jennings (19951) for a survey). This Third, how do we ensure that agents act coherently axiomatization is accomplished by formalizing a others, or in making decisions or taking action, accommodating model for agent behavior in terms of beliefs, desires, explain and the non-local effects of local decisions and avoiding goals, and so on- These works are known as belief- harmful interactions? How do we ensure the MAS does desire-intention (BDI) systems (see Rao and Georgeff possibly not become resource bounded? How do we [1991] and Shoham (19931). An agent that has a BDI- repair conflicts avoid unstable system behavior? type architecture has also been called deliberative. Fourth, how do we enable individual agents to In my own work on the RETSINA multiagent and harmful represent and reason about the actions, plans, infrastructure, agents coordinate to gather information interactions. and knowledge of other agents to coordinate with them; in the context of user problem-solving tasks. Each how do we reason about the state of their coordinated RETSINA agent is a BDI-type agent that integrates process (for example, initiation and completion)? planning, scheduling, execution, information Fifth, how do we recognize and reconcile disparate gathering, and coordination with other agents (Decker, viewpoints and conflicting intentions among a Pannu, et al. 1997; Sycara et at. 1996). Each agent has collection of agents trying to coordinate their actions? a sophisticated reasoning architecture that consists of Sixth, how do we engineer and constrain practical different modules that operate asynchronously. DAI systems? How do we design technology platforms The planning module takes as input a set of goals and and development methodologies for MASs? produces a plan that satisfies the goals. The agent Solutions to these problems are intertwined (Gasser planning process is based on a hierarchical task 1991). For example, different modeling schemes of an network (HTN) planning formalism. It takes as input individual agent can constrain the range of effective the agents current set of goals, the current set of task coordination regimes; different procedures for structures, and a library of task-reduction schemas. A communication and interaction have implications for task-reduction schema presents a way of carrying out a behavioral coherence. Different problem and task task by specifying a set of subtasks-actions and decompositions can yield different interactions. It is describing the information-flow relationships between arguable whether one can find a unique most important them. (See Williamson, Decker, and Sycara [19961]). dimension along which a treatment of MASs can The communication and coordination module accepts cogently be organized. Here, I attempt to use the dimension of effective overall problem-solving and interprets messages coherence of an MAS as the organizing theme.

characteristic is a strength of the approach because the from other agents in KQML. Messages can contain requests for services. agents do not need to revise their world model as it Reactive agents These requests become goals of the recipient agent. The scheduling changes. Thus, robustness and fault tolerance are two do not have module schedules each of the plan steps. The agent scheduling process of the main properties of reactive systems. A group of representations takes as input, in general, the agents current set of plan instances and, in agents can complete a task even if one of them fails. particular, the set of all executable actions and decides which action, if However, purely reactive systems suffer from two of their main limitations: any, is to be executed next This action is then identified as a fixed environment intention until it is actually carried out (by the execution component). First, because purely reactive agents make decisions Agent-reactivity considerations are handled by the execution-monitoring based on local information, they cannot take into and act using a process. Execution monitoring takes as input the agents next intended consideration non-local information or predict the stimulus- action and prepares, monitors, and completes its execution. The effect of their decisions on global behavior Such execution monitor prepares an action for execution by setting up a myopic behavior could lead to unpredictable and response type context (including the results of previous actions) for the action. It unstable system behavior (Thomas and Sycara 1998; of behavior; monitors the action by optionally providing the associated computation- Huberman and Hogg 1988). In reactive systems, the they respond to limited resources--for example, the action might be allowed only a relationship between individual behaviors, environ- certain amount of time, and if the action does not complete before the ment, and overall behavior is not understandable, the present which necessarily makes it hard to engineer agents to time is up, the computation is interrupted, and the action is marked as fulfill specific tasks: One must use a laborious process state of the having failed. Failed actions are handled by the exception-handling of experimentation, trial, and error to engineer an agent process. The agent has a domain-independent library of plan fragments environment in or an MAS. Despite these disadvantages, reactive sys- (task structures) that are indexed by goals as well as a domain-specific tems have advantages of speed (the sophisticated which they are library of plan fragments that can be retrieved and incrementally reasoning of deliberative agents can slow them; hence, instantiated according to the current input parameters. situated. they are useful in rapidly changing environments). Reactive agents have also been developed. Reactive agents have their In fact, for most problems, neither a purely deliberative roots in Brooks’s (1991) criticism of deliberative agents and his nor a purely reactive architecture is appropriate, but assertions that (1) intelligence is the product of the interaction of an hybrid architectures can combine aspects of both. agent and its environment and (2) intelligent behavior emerges from the Typically, these architectures are realized as a number interaction of various simpler behaviors organized in a layered way of software layers, each dealing with a different level through a master-slave relationship of inhibition (subsumption archi- of abstraction. Most architectures find three layers tecture). A different reactive architecture is based on considering the sufficient. Thus, at the lowest level in the hierarchy, behavior of an agent as the result of competing entities trying to get there is typically a reactive layer, which makes control over the actions of the agent This idea has its root in the society decisions about what to do based on raw sensor input of the mind (Minsky 1986) and has been employed in different ways by The middle layer typically abstracts away from raw Maes (1990), Travers (1988), and Drogul and Ferber (Ferber 1996; sensor input and deals with a knowledge-level view of Drogul and Ferber 1992). An agent is defined as a set of conflicting tasks the agents environment, typically making use of where only one can be active simultaneously. A task is a high-level symbolic representations. The uppermost level of the behavioral sequence as opposed to the low-level actions performed architecture tends to deal with the social aspects of the directly by actuators. A reinforcement mechanism is used as a basic environment. Coordination with other agents is learning tool to allow the agents to learn to be more efficient in tasks that typically represented in the uppermost layer .The way are often used. This architecture has been used in the MANTA system to that the layers interact with one another to produce the simulate the behavior of ants (Ferber 1996). global behavior of the agent differs from architecture Reactive agents do not have representations of their environment and to architecture (Bonasso et at. 1996; Ferguson 1995, act using a stimulus response type of behavior they respond to the 1992; Muller and Pischel 1994). present state of the environment in which they are situated. They do not The area of agent architectures, particularly layered take history into account or plan for the future. Through simple architectures, continues to be an area interactions with other agents, complex global behavior can emerge. This

82 SUMMER 1998 Articles

of considerable research effort within the Another perspective in DAI defines organiza- how a pluralistic community could operate multiagent field. For example, there is ongoing tion less in terms of structure and more in (Kornfeld and Hewitt 1981). Solutions to work to investigate the appropriateness of var- terms of current organization theory. For problems are locally constructed, then they are ious architectures for different environment example, Gasser (1986) views an organization communicated to other problem solvers that types. It turns out to be hard to evaluate one as a "particular set of settled and unsettled can test, challenge, and refine the solution agent architecture against another, especially questions about beliefs and actions through (Lesser 1991).

within the context of an MAS. which agents view other agents." In this view, In open, dynamic environments, the issue an organization is defined as a set of agents of organizational adaptivity is crucial. Organizations with mutual commitments, global commit- Organizations that can adapt to changing An organization provides a framework for agent ments, and mutual beliefs (Bond and Gasser circumstances by altering the pattern of interactions through the definition of roles, 1988). An organization consists of a group of interactions among the different constituent behavior expectations, and authority relations. agents, a set of activities performed by the agents have the potential to achieve coherence Organizations are, in general, conceptualized in agents, a set of connections among agents, and in changing and open environments. RETSINA terms of their structure, that is, the pattern of a set of goals or evaluation criteria by which exhibits organizational adaptivity through information and control relations that exist the combined activities of the agents are eval- cooperation mediated by middle agents. among agents and the distribution of problem - uated. The organizational structure imposes RETSINA agents find their collaborators solving capabilities among them. In cooperative constraints on the ways the agents dynamically based on the requirements of the problem solving, for example (Corkill and communicate and coordinate. Examples of task and on which agents are part of the Lesser 1983), a structure gives each agent a high organizations that have been explored in the society at any given time, thus adaptively level view of how the group solves problems. MAS literature include the following: forming teams on demand (Decker, The structure should also indicate the connec- Hierarchy: The authority for decision- Williamson, and Sycara 1996a). tivity information to the agents so they can making and control is concentrated in a single distribute subproblems to competent agents. problem solver (or specialized group) at each In open-world environments, agents in the level in the hierarchy. Interaction is through system are not statically predefined but can vertical communication from superior to sub- Organizations are, in general, dynamically enter and exit an organization, ordinate agent, and vice versa. Superior agents which necessitates mechanisms for agent locat- exercise control over resources and decision- conceptualized in terms of ing. This task is challenging, especially in envi- making. their structure, that is, the ronments that include large numbers of agents Community of experts: This organization is and that have information sources, communi- flat, where each problem solver is a specialist pattern of information cation links, and/or agents that might be in some particular area. The agents interact by and control relations appearing and disappearing. Researchers have rules of order and behavior (Lewis and Sycara identified different kinds of middle agent 1993; Lander, Lesser, and Connell 1991). that exist among agents and (Decker, Sycara, and Williamson 1997) that help Agents coordinate though mutual adjustment the distribution of problem- agents find others. When an agent is of their solutions so that overall coherence can solving capabilities among instantiated, it advertises its capabilities to a be achieved. middle agent An agent that is looking to find Market: Control is distributed to the agents them. another that possesses a particular capability (for that compete for tasks or resources through example, can supply particular information or bidding and contractual mechanisms. Agents Task Allocation achieve a problem-solving goal) can query a interact through one variable, price, which is Task allocation is the problem of assigning middle agent In the RETSINA infrastructure, used to value services (Mullen and Wellman responsibility and problem-solving resources there could be multiple middle agents, not only 1996; Davis and Smith 1983; Sandholm 1993). to an agent. Minimizing task interdependen- in type but also in number For example, Agents coordinate through mutual adjustment cies has two general benefits regarding coher- protocols have been developed for distributed of prices. ence: First, it improves problem-solving effi- matchmaking (Jha et at. 1998). Scientific community: This is a model of ciency by decreasing communication overhead

SUMMER 1998

among the problem-solving agents. Second, Lesser (1995), where decommitment approach to resolving subproblem it improves the chances for solution penalties were introduced, and by Sycara interdependencies that result from an consistency by minimizing potential (1997), where the theory of financial option incomplete set of data for each agent that conflicts. In the second case, it also improves pricing has been used to achieve flexible could allow it to solve a subproblem tern efficiency because resolving conflicts can be contracting schemes in uncertain completely is the a time-consuming process. environment. The issue of task allocation was one of FUNCTIONALLY ACCURATE MODEL (FA/C) the earliest problems to be worked on in DAI Multiagent Planning research. On the one extreme, the designer (Lesser 1991). In the PA/C model, agents do can make all the task assignments in Agents can improve coherence by planning not need to have all the necessary information advance, thus creating a nonadaptive their actions. Planning for a single agent is a locally to solve their subproblems but instead problem-solving organization. This approach process of constructing a sequence of actions interact through the asynchronous, coroutine is limiting and inflexible for environments considering only goals, capabilities, and exchange of partial results. With the FA/C with a high degree of dynamism, openness, environmental constraints. However, plan- model, a series of sophisticated distributed and uncertainty. However, one can do task ning in an MAS environment also considers control schemes for agent coordination were allocation dynamically and flexibly. The the constraints that the other agents’ activities developed, such as the use of static metalevel issue of flexible allocation of tasks to place on an agent’s choice of actions, the information specified by an organizational multiple agents received attention early on constraints that an agent’s commitments to structure and the use of dynamic metalevel (Davis and Smith 1983). Davis and Smith’s others place on its own choice of actions, information developed in partial global work resulted in the well-known and the unpredictable evolution of the world planning (PGP) (Durfee 1987). CONTRACT NET PROTOCOL (CNP). In caused by other unmodeled agents. POP is a flexible approach to coordination this protocol, agents can dynamically take Most early work in DAI has dealt with that does not assume any particular two roles: (1) manager or (2) contractor. groups of agents pursuing common goals distribution of subproblems, expertise, or Given a task to perform, an agent first deter- (for example, Lesser [1991); Lesser, Durfee, other resources but, instead, lets nodes mines whether it can break it into sub-tasks and Corkill [1989]; Durfee [19881; Conry, coordinate in response to current situations that can be performed concurrently. It Meyer and Lesser [19881; Cammaarata, (Durfee 1987)- Agent interactions take the utilizes the protocol to announce the tasks McArthur, and Steeb [19831). Agent form of communicating plans and goals at an that could be transferred and requests bids interactions are guided by cooperation appropriate level of abstraction. These from agents that could perform any of these strategies meant to improve their collective communications enable a receiving agent to tasks. An agent that receives a task- performance. Most work on multiagent form expectations about the future behavior announcement message replies with a bid for planning assumes an individual sophisticated of a sending agent, thus improving agent the task, indicating how well it thinks it can agent architecture that enables them to do predictability and network coherence (Durfee perform the task. The contractor collects the rather complex reasoning. Early work on 1988). Because agents are cooperative, the bids and awards the task to the best bidder. distributed planning took the approach of recipient agent uses the information in the The CNP allows nodes to broadcast bid complete planning before action. To produce plan to adjust its own local planning appro- requests to all others. The CNP was sub- a coherent plan, the agents must be able to priately, so that the common planning goals sequently used to control factory floor recognize subgoal interactions and avoid (and planning effectiveness criteria) are met operations (Parunak 1987). AIthough CNP them or resolve them. Early work by Besides their common POPs, agents also was considered by Smith and Davis and Georgeff (Rao and Georgeff 1991) included have some common knowledge about how many DAI researchers to be a negotiation a static agent to recognize and resolve such and when to use POPs. Decker and Lesser principle, it is a coordination method for task interactions. The agents sent this agent their (1995) addressed some of the limitations of allocation, CNP enables dynamic task plan; the agent examined them for critical the POP by creating a generic POP-based allocation, allows agents to bid for multiple regions where, for example, contention for framework called TAEMS to handle issues of tasks at a time, and provides natural load resources could cause them to fail. The agent real time (for example, scheduling to balancing (busy agents need not bid). Its then inserted synchronization messages (akin deadlines) and metacontrol (for example, to limitations are that It does not detect or to operating system semaphores) so that one obviate the need to do detailed planning at all resolve conflicts, the manager does not agent would wait till the resource was possible node interactions). TAUV45 has been inform nodes whose bids have been refused, released by another agent In the work by used as a framework for evaluation of agents cannot refuse bids, there is no Cammarata on air-traffic control (Steeb et al. coordination algorithms. preemption in task execution (time-critical 1988; Cammarata, McArthur, and Steeb Another direction of research in tasks might not be attended to), and it is 1983), the synchronizing agent was cooperative multiagent planning has communication intensive. Extensions dynamically assigned according to different to CNP have been made by Sandholm and criteria, and it could alter its p1w to remove the interaction (avoid collision). Another

84 been focused on modeling teamwork about 300 domain-independent SOAR rules. reasoning and argumentation, where it might explicitly. Explicit modeling of teamwork is Based on the teamwork operationalized in be possible to construct arguments in support particularly helpful in dynamic environments STEAM, three applications have been of a particular perspective (Loin 1987; where team members might fail or be implemented, two that operate in a Hewitt 1986, Kornfeld and Hewitt 1981; presented with new opportunities. In such commercially available simulation for Lesser and Corkill 1981), or it might be situations, it is necessary that teams monitor military training and the third that is part of possible to construct persuasive arguments to their performance and reorganize based on RODOCLJP synthetic soccer. STEAM uses a change the intentions, beliefs, preferences, the situation. hybrid approach that combines joint and actions of a persuadee (Sycara 1990b) so The joint-intentions framework (Cohen intentions (Cohen and Levesque 1990) but that effective plans and solutions can be and Levesque 1990) focuses on also uses partial SHAREDPLANS (Grosz produced; (3) constraint relaxation, where characterizing a team’s mental state, called a and Kraus 1996). conflicting constraints can be resolved by joint intention. A team jointly intends a team Increasingly, the emphasis of multi-agent relaxing them (Liu and Sycara 1997; Sycara action if the ream members are jointly planning has been on flexible et al. 1991) or reformulating the problem to committed to completing the team action communication and action execution in eliminate the constraints (Sycara 1991); and while mutually believing they were doing it. complex, dynamic environments, including (4) social norms that impose some sort of A joint commitment is defined as a joint agents that might be hostile or at least self- common standard of behavior, which when persistent goal. To enter into a joint interested (Veloso et al. 1997) and perform adhered to can lead to conflict avoidance commitment, all team members must well in dynamically changing environments (Castelfranchi, Miceli and Cesta 1992). establish appropriate mutual beliefs and (Kinney et al. 1992). Social norms and standards, although commitments, which is done through an helping avoid conflicts, can impede exchange of request and confirm speech acts Recognizing and adaptation. (Cohen and Levesque 1990). The Resolving Conflicts Negotiation is seen as a method for commitment protocol synchronizes the team Because MASs lack global viewpoints, coordination and conflict resolution (for in that all members simultaneously enter into global knowledge, and global control, there example, resolving goal disparities in a joint commitment toward a team task. In is the potential for disparities and planning, resolving constraints in resource addition, all team members must consent, inconsistencies in agents’ goals, plans, allocation, resolving task inconsistencies in using confirmation, to the establishment of a knowledge, beliefs, and results. To achieve determining organizational structure). joint commitment goal. If a team member coherent problem solving, these disparities Negotiation has also been used as a refuses, negotiation could be used; however, must be recognized and resolved. Disparities metaphor for communication of plan how it is done remains an open issue. can be resolved by making an agent omni- changes, task allocation (for example, CNP), The model of SHAREDPLAN (Own and scient so it can see the states of all agents or centralized resolution of constraint Kraus 1996; Grosz and Sidner 1990) is not and determine where the disparities lie and violations. Hence, negotiation is almost as ill based on a joint mental attitude but rather on how to resolve them. This approach is defined as the notion of agent. I give here a new mental attitude intending that an action limiting because it makes this agent a what I consider as the main characteristics of be done. Intending is defined using a set of bottleneck and a single point of failure. To negotiation that are necessary for developing axioms that guide a teammate to take action detect and correct disparities and conflicts applications in the real world (1) the using only local perspective is difficult To or enter into communication that enables or presence of some sort of conflict that must facilitate detection and resolution of facilitates its teammates to perform assigned be resolved in a decentralized manner by (2) conflicts, agents can rely on models of the tasks. COLLAGEN (Rick and Sidner 1997) is a self-interested agents under conditions of (3) world and other agents- Disparity resolution prototype tooikit, which has its origins in bounded rationality and (4) incomplete can be influenced by the organizational SHAREDPLAN, and is applied to building a information. Furthermore, the agents structure of the agent society and an agent’s collaborative interface agent that helps with communicate and iteratively exchange role within it, the kinds of models an agent air-travel arrangements. Jennings (1995) proposals and counterproposals. has, and the agent’s reasoning algorithms. presented a framework called joint My PERSUADER system (1990b, 1988, The main approach for resolving disparities responsibility based on a joint commitment to 1987) and work by Rosenschein in an MAS is negotiation. Before I treat (Rosenschein and Zlotkin 1994; Rosenschein a team’s joint goal and a joint recipe commit- negotiation in some detail, I present some and Genesereth 1985; Zlotkln and ment to a common recipe. This model was less often used approaches that are organized Rosenschein 1991) are the first AI research Implemented in the ORATE system. in Gasser (1992). They include (1) on negotiation among self-interested agents. Tambe (1997) presents a model of teamwork, assumption surfacing, where inconsistent The two approaches differ in their assump- called STEAM (a shell for TEAMWORK), propositions can be backed up to their tions, motivations, and operationalism. based on enhancements to the SOAR assumptions (Huhns and Bridgeland 1991; architecture (Newell 1990), plus a set of Mason and Johnson 1989); (2) evidential

85 The work of Rosenschein was based on game the negotiation. what to communicate or allowing theory. Utility is the single issue that the Liu and Sycara (1994) and Garrido and coordination without communication. The agents consider, and the agents are assumed Sycara (1996) have modeled negotiation as a ability to model others increases an agent’s omniscient. Utility values for alternative constraint-relaxation process where the flexibility. Instead of operating on the basis outcomes are represented in a payoff matrix agents are self-interested in the sense that of a fixed protocol of interaction, modeling that is common knowledge to both parties in they would like to achieve an agreement that others allows the agent to change the pattern the negotiation. Each party reasons about and gives them the highest utility but are also of interaction. Having a model allows an chooses the alternative that will maximize its cooperative in the sense that they are willing agent to infer events or behaviors of the utility. Despite the mathematical elegance of to accept lower utility to facilitate reaching other agent that it cannot sense directly. game theory, game-theoretic models suffer an agreement. The agents communicate their Several typical components of agent models from restrictive assumptions that limit their constraints through proposals and are commitments, capabilities, resources, applicability to realistic problems.1 Real- counterproposals. The application domain of beliefs, plans, and organizational world negotiations are conducted under the work has been a distributed meeting knowledge. We saw already how explicit uncertainty they involve multiple criteria scheduling. In Garrido and Sycara (1996), commitments to joint activity are the rather than a single utility dimension, the experimental results were obtained to study cornerstone of models of teamwork. In my utilities of the agents are not common the scheduling efficiency and quality of the own work on the RETSINA multiagent knowledge but private, and the agents are not agreed-on final schedule under various Infrastructure, commitment of an agent to omniscient. conditions of information privacy and pref- performing a task is not communicated PER5UAOER (Sycara 1990a) is an erence structures of the participants. explicitly to others but is implicit in its implemented system that involves three Because electronic commerce is rapidly advertisement to a middle agent. An agent’s agents--1) a union, (2) a company, and (3) a becoming a reality, the need for negotiation advertisement describes its capabilities, that mediator--and it operates in the domain of techniques that take into consideration the is, the set of services it can provide to others. labor negotiation. It is inspired by human complexities of the real world, such as For example, the advertisement of an negotiation. It models iterative exchange of incomplete information, multiple negotiation information agent expresses the set of proposals and counterproposals for the parties issues, negotiation deadlines, and the ability queries that the agent is capable of to reach agreement The negotiation involves to break contracts, will critically be needed. answering. In this way, an agent that needs a multiple issues, such as wages, pensions, Work in nonbinding contracts includes particular service can, through asking a seniority, and subcontracting. Each agent’s Sandholrn and Lesser (1995), where middle agent, find out who is capable of multidimensional utility model is private decommitment penalties were introduced providing the service and can contact the (rather than common) knowledge. Belief into the CNP, and my work (Sycara 1997), service provider(s). Thus, an agent’s revision to change the agents' utilities so that where financial option pricing has been used advertisement is the model of itself it makes agreement can be reached is done through to model the value of a contingent contract available to the agent society. Thus, for persuasive argumentation (Sycara 1990b). In and calculate optimal values of different example, Instead of agent X either having a addition, case-based learning is also contracting parameters of a contingent model of agent Y built into It at design time incorporated in the model. The negotiation contract, such as when to give a contract to a or having to infer it, agent Y makes its model of PER5UADEIL has also been applied capability model available to agent X contractee, when to break a contract, and to the domain of concurrent engineering through an advertisement to a middle agent which contract to accept of a set of offered (Lewis and Sycara 1993). Parsons and These two ways would be infeasible and con-tracts. Jennings (1996) have followed the formalism extremely time consuming for agents in described in Kraus, Nirke, and Sycara (1993) Modeling Other Agents large-scale and open environments such as to construct arguments to evaluate proposals the internet. Providing a model of oneself, Agents can increase the accuracy and and counterproposals in negotiation. however, is a solution that is scalable and efficiency of their problem solving if they are Work by Kraus, Wilkenfeld, and Zlotkin works in an open environment. If the agent given knowledge about other agents or the (1995) focuses on the role of time in goes down, it "unadvertises"; that is, it lets it ability to model and reason about others. negotiation. Using a distributed mechanism, be known that it is no longer a member of They can utilize this knowledge to predict the agents conduct negotiation and can reach the society. possible conflicts and interpret, explain, and efficient agreements without delays. It is also predict the other agents’ actions under shown that the individual approach of each Managing Communication varying circumstances. In addition, modeling agent toward the negotiation time affects Agents can improve the coherence of their others provides information to an and even determines the final agreement that problem solving by planning the content, agent about others' knowledge, beliefs, and is reached. The main conclusion is that amount, type, and timing of goals that are useful in planning when and delaying agreements causes inefficiency in

the communication they exchange. It has been (Finin et al. 1994) and the one results on standard benchmark problems from noted (Durfee, Lesser, and Corkil1 987) that the operations research literature in problem using abstraction and metalevel information Economics-based approaches, such as market (for example, organizational knowledge) is helpful because they help decrease mechanisms, are becoming increasingly attractive communication overhead, in dynamic and to MAS researchers because of the ready open environments, inhabited by heterogeneous agents, additional issues need to availability of underlying formal models and their be faced. The most prominent among them is potential applicability in internet-based commerce. agent interoperability. Agents are populating the internet at a developed by Cohen and his colleagues (Smith scalability (as many as 5000 agents), rapid pace. These agents have different and Cohen 1996) that provide a set of computational efficiency, and high-solution functions (capabilities). There can, however, performatives based on speech acts, have been quality under different optimization criteria be many agents with the same functions (for designed. Although such performatives can (for example, the minimization of makespan, example, many information agents provide characterize message types, efficient languages weighted tardiness, and so on) (flu and financial news). Agents can increase their to express message content that allows agents to Sycara problem-solving scope by cooperation. In an understand each other have not been 1993). open environment, heterogeneous agents that demonstrated effectively (although KIF Economic-based mechanisms have been would like to coordinate with each other (Genesereth and Ketchpel 1994] has been pro- utilized in MASs to address problems of (either cooperate or negotiate, for example) posed). The ontology problem, that is, how resource allocation (the central theme of face two major challenges: First, they must be agents can share meaning, is still open (Gruber economic research). Economics-based able to find each other (in an open 1993). approaches, such as market mechanisms, are environment, agents might appear and becoming increasingly attractive to MAS re- Managing Resources disappear unpredictably), and second, they searchers because of the ready avail-ability of Another critical issue is effective allocation of must be able to inter-operate. underlying formal models and their potential limited resources to multiple agents. For To address the issue of finding agents in an applicability in internet-based commerce. In example we have all experienced large time open environment such as the internet, middle economics-based approaches, agents are lags in response to - -internet queries because of agents (Decker, Williamson, and Sycara assumed to be self-interested utility network congestion. 1996b) have been proposed. Different agent maximizers. In markers, agents that control Various approaches have been developed for types were identified and implemented scarce resources (labor, raw materials, goods, effective resource allocation to multiple agents. money) agree to share by exchanging some of (Decker, Sycara, and Williamson 1997). Some of them hail from operations research- These types include matchmakers or yellow their respective resources to achieve some based techniques for single-agent scheduling, page agents that match advertisements to common goal. Resources are a-changed with and others use market-oriented approaches. In requests for advertised capabilities, or without explicit prices. Market the first category of approaches, I mention blackboard agents that collect requests, and mechanisms have been used for resource distributed constraint heuristic search (Sycara brokers that-process both. In preliminary allocation (Mullen and Wellman 1996; et al. 1991), which combines decentralized - experiments (Decker, Sycara, and Williamson Huberman and Clearwater 1995; Sandholm search with constraint satisfaction and 1997), it was seen that the behaviors of each 1993). Markets assume that the exchange optimization. The method relies on (1) a set of type of middle agent have certain performance prices are publicly known. In auctions, there variable and value-ordering heuristics that characteristics; for example, although is a central auctioneer through which quantify several characteristics of the space brokered systems are more vulnerable to coordination happens. Hence, the agents being searched and (2) a communication certain failures, they are also able to cope exchange minimal amounts of information. protocol that allows the agents to coordinate in more quickly with a rapidly fluctuating agent It is probable that In the future, most an effective manner. Another distributed work force. Middle agents are advantageous agents will be self-interested. A self- scheduling model exploits a large number of because they allow a system to operate interested agent simply chooses a course of simple agents by partitioning problem robustly in the face of agent appearance and action that maximizes its own utility In a constraints and assigning them to specialized disappearance and intermittent communica- society of self-interested agents It is desired agent classes- This methodology was applied to tions. that if each agent maximizes its local utility, solve job-shop-scheduling constraint-satisfac- the whole society exhibits good behavior, in To allow agents to interoperate, tion and constraint-optimization problems (Liu other words, good local behavior Implies communication languages, such as KQML and Sycara 1997, 1995a, 1995b) with good good global behavior. The goal

SUMMER 1998

87 is to design mechanisms for self-inter- the environment effectively changes learn, the joint system utility is near ested agents such that if agents follow as other agents learn. Moreover, other optimal, and agents’ individual utili- these mechanisms, the overall system agents’ actions are often not directly ties are similar (2) when no agent behavior will be acceptable, which is observable, and the action taken by learns, the agents’ individual utilities called mechanism design. Many prob- the learning agent can strongly bias are almost equal, but the joint utility is ems face such a society of self-inter- which range of behaviors is encoun- low (much lower than in the all- esred agents: First, agents might tered. In Hu and Wellman (1996), an agents- learn condition); and (3) when overuse and, hence, congest a shared agent’s belief process is characterized only one agent learns, its individual resource, such as a communications in terms of conjectures about the utility increases at the expense of both network. This problem is called the effect of its actions. A conjectural equi- the individual utility of the other tragedy of the commons (Hardin 1968). librium is then defined where all agents as well as the overall joint util- Generally, the problem of tragedy of agents’ expectations are realized, and ity of the system (that is, only one commons is solved by pricing or tax- each agent responds optimally to its agent learning has a harmful overall ing schemes. expectations. An MAS is presented effect) (Zeng and Sycara 1998,1997). Second, a society of self-interested where an agent builds a model of the A survey of multiagent learning can be computational agents can exhibit response of others. Reported experi- found in Stone and Veloso (1997). oscillatory or chaotic behavior mental results show that depending (Thomas and Sycara 1998; Huberman on the starting point, an agent might Multiagent System Applications and Hogg 1988). The experimental be better or worse off than had it not results show that imperfect knowledge attempted to learn a model of the oth- suppresses oscillatory behavior at the er agents. Such equivocal results have The first MAS applications appeared in expense of reducing performance. also been observed in computational the mid-1980s and increasingly cover Moreover, systems can remain in ecosystems (Huberman and Hogg a variety of domains, ranging from nonoptimal metastable states for 1988), where a large number of simple manufacturing to process control, air- extremely long times before getting to agents operate asynchronously with- traffic control, and information man- the globally optimal state. In Thomas out central control and compete for agement. I give here a few representa- and Sycara (1998), a similar problem is resources, with incomplete knowledge tive examples. For a more detailed considered. Two approaches are evalu- and delayed information. Kephart, description of agent-based applica- ated: (1) heterogeneous preferences Hogg. and Huberman (1989) per- tions, see Jennings et al. (1998) and and (2) heterogeneous transaction formed experimental nut ions and C1,-draa (1995). costs; empirically, the transaction cost verified some of the previous theoreti- One of the earliest MAS applications case provides stability with near-opti- cal results; thus, for certain parameter was distributed vehicle monitoring mal payoffs under certain conditions, settings, these systems exhibit behav- (DVMT) (Durfee 1996, 1987; Durfee Third, agents might be untruthful ior characterized by fixed points, osci- and Lesser 1989), where a set of geo- or deceitful to increase their individ- lation, and even chaos. Complex graphically distributed agents monitor ual utility. Lying and deceitfulness can behavior thus cube exhibited by is- vehicles that pass through their have harmful effects on the whole pie computational ecosystems. En- respective areas, attempt to come up society. Mechanism design techniques hancing the decision-making abilities with interpretations of what vehicles have been reported that make it bene- of some of the individuals in the are passing through the global area, ficial for agents to report the truth I system can either improve or severely and track vehicle movements. The (Myerson 1989). Another body of degrade overall system performance. DVMT has been used as an MAS test research explores the effective alloca- Recently, there has been increasing bed. tion of computational resources and interest in integrating, learning into Parunak (1987) describes the YAMS load-balancing issues. In Shehory, Thai the negotiation process (for example, (YET ANOTHER MANUFACTURING SYSTEM) and Sycara (1997). agent-cloning and Sen (1996). Zeng and Sycara (1997) system, which applies the CNP to man- agent-merging techniques were devel- have developed an economic bargain- ufacturing control. The basic problem oped as a way to mitigate agent over- ing negotiation model, BAZAR, where can be described as follows: A manu- loading and promote system load balancing. the agents are self-interested. The facturing enterprise is modeled as a model emphasizes the learning hierarchy of functionally specific work aspects. The agents keep histories of cells. These work cells are further Adaptation and Learning their interactions and update their grouped into flexible manufacturing MAS adaptivity to changing circum- beliefs, using Bayesian updating, after systems (FMSS) that collectively consti- stances by altering the problem-solv- observing their environment and the tute a factory. The goal of YAMS is to ing behavior of Individual agents or behavior of the other negotiating efficiently manage the production the patterns of agent interactions pro- agents. The benefits of learning, if any, process of these factories. To achieve vides the potential for increasing on the individual utilities of agents, as this complex task, YAMS adopts a mul- problem-solving coherence. well as the overall (joint) system utili- tiagent approach, where each factory Learning In a multiagent environ- ty, were examined. The experimental and factory component is represented ment is complicated by the fact that results suggest that (1) when all agents as an agent. Each agent has a collec

AI MAGAZINE

Articles portfolio. The system consists of agents that needed that enable designers to specify an tion of plans representing its capabilities. The cooperatively self-organize to monitor and agent’s problem-solving behavior, specify CNP allows tasks (that is, production orders) track stock quotes, financial news, financial how and when agents should interact, and to be delegated to Individual factories, and analyst reports, and company earnings visualize and debug the problem-solving from individual factories down to FMSs, and reports to appraise the portfolio owner of behavior of the agents and the entire system. then to individual work cells. Other systems in the evolving financial picture. The agents The other major impediment to the this area include those for configuration not only answer relevant queries but also widespread adoption of agent technology has design of manufacturing products (Dan and continuously monitor internet information a social, as well as a technical, aspect. For Birmingham 1996) and collaborative design sources for the occurrence of interesting individuals to be comfortable with the idea of (Cutkosky et al. events (for example, a particular stock has delegating tasks to agents, they must first 1993). gone up past a threshold) and alert the trust them (Bradshaw 1997; Maes 1994)- The The best-known MAS for process control is portfolio manager agent or the user. process of mutual adjustment between user ARCHON, a software platform for building WARREN also includes agents that analyze and agents (both in terms of the agent MASs and an associated methodology for user buy and sell decisions with respect to learning user preferences but also in terms of building applications with this platform asset allocations and risk (Sycara, Decker, the user learning agents’ capabilities and (Jennings, Corera, and Laresgoiti 1995). and Zeng 1998). limitations) takes time. During this period, Archon has been applied in several process- In addition, there are a variety of MAS agents must strike a balance between control applications, including electricity applications in telecommunications. In one continually seeking guidance (and needlessly transportation management (the application is such application (Weihmayer and distracting the user) and never seeking in use in northern Spain) and particle Velthuijsen 1994), Griffeth and Velthuijsen guidance (and exceeding their authority). accelerator control. ARCHON also has the use negotiating agents to tackle the feature distinction of being one of the world’s earliest interaction problem by utilizing negotiating Acknowledgments field-tested MASs. Other agent-based process- agents to represent the different entities that This work on agent-based systems has been control systems have been written for are interested in the set up of a call. When supported by Office of Naval Research grant monitoring and diagnosing faults in nuclear conflicts are detected, the agents negotiate N-00014-96-1-1222, Advanced Research power plants (Wang and Wang 1997), with one another to resolve them so that an Projects Agency contract F33615-93-1-1330, spacecraft control (Schwuttke and Quan acceptable call configuration can be and National Science Foundation grant IRI- 1993), and climate control (Huberman and established. Other problems for which 9612131. Clearwater 1995). agent-based systems have been constructed Notes Ljunberg and Lucas (1992) describe a include network control, transmission and 1. It should be noted that some of the very sophisticated agent-based air-traffic control switching, service management, and recent game-theoretic models are directly system known as OASIS. In this system, network management motivated by considerations of dropping or which is undergoing field trials at the Sydney, relaxing some of these assumptions. Although Australia, airport, agents are used to represent there has been interesting progress reported both aircraft and the various air-traffic control Conclusions in the literature (for example, Jordan (1992]), the systems in operation. The agent metaphor thus Designing and building agent systems is fundamental framework and methodology of provides a useful and natural way of modeling difficult. They have all the problems game theory remains almost the same, and it real-world autonomous components. As an associated with building traditional might be too early to tell whether these new aircraft enters Sydney airspace, an agent is distributed, concurrent systems and have the results will reshape the current game-theo- allocated for it, and the agent is instantiated additional difficulties that arise from having retic framework. 2. Existing technologies such as CORBA are at a with the information and goals corresponding flexible and sophisticated interactions low level and, thus, unable to provide the support to the real-world aircraft. For example, an between autonomous problem-solving needed for the structuring of flexible multiagent aircraft might have a goal to land on a certain components. The big question then becomes systems. runway at a certain time. Air-traffic control one of how effective MASs can be designed agents are responsible for managing the and implemented. References system. OASIS is implemented using the At this time, there are two major Bonasso, IL R; Kortenkamp, D.; Miller, D. Australian Artificial Intelligence Institute’s technical impediments to the widespread P; and Slack, M. 1996. Experiences with an own BDI model of agency (DMARS). adoption of multiagent technology: (1) the Architecture for Intelligent, Reactive Agents. The WARREN financial portfolio man- lack of a systematic methodology enabling In Intelligent Agents II, eds. M. Wooldridge. agement system (Sycara et al. 1996) is an designers to clearly specify and structure J. P. Muller, and M. Tambe.187-202. Lecture MAS that integrates information finding and their applications as MASs and (2) the lack Notes In Artificial Intelligence 1037. New York filtering from the internet in the context of of widely available industrial-strength Springer-Verlag. supporting a user manage his/her financial MAS toolkits.2 Flexible sets of tools are

SUMMER 1998

89 Bond, A. H., and Gasser, I.. 1988. Readings ceedings of the First International Conference on Ferber, J. 1996. Reactive Distributed Artifi- in Distributed Artificial Intelligence. San Fran- Multiagent Systems, 73-80. Menlo Park, Calif.: cial Intelligence: Principles and Applications. In cisco, Calif.: Morgan Kaufmann. AAAI Press. Foundations of Distributed Artificial Intelligence, Bradshaw, J. 1997. Software Agents. Menlo Decker, K.; Pannu. A.; Sycara, K; and eds. G. 0-Hare and N. Jennings, 287-314. New York Park. Calif.: AAAI Press. Williamson, M. 1997. Designing Behaviors for Wiley. Brooks, R. A. 1991. Intelligence without Information Agents. In Proceedings of the Ferguson. I. A. 1995. Integrated Control and Representation. Artificial Intelligence 47(1-3): First International Conference on Autonomous Coordinated Behavior: A Case for Agent 139-159. Agents (Agent-97), 404-412. New York: Models. In Intelligent Agents: Theories, Cammarata. S.; McArthur, D.; and Steeb, Association of Computing Machinery. Architectures, and Languages, eds. M. Wooldridge

II. 1983. Strategies of Cooperation in Dis- Decker, K; Sycara, K, and Williamson, M. and N. K Jennings, 203-218. Lecture Notes in tributed Problem Solving. In Proceedings of 1997. Middle Agents for the Internet. In Artificial Intelligence, Volume 890. New York. the Eighth International Joint Conference on Proceedings of the Fifteenth International Joint Springer-Verlag. Artificial Intelligence (IJCAI-83), 767-770. Conference on Artificial Intelligence (IJCAI-97), Ferguson, I. A. 1992. TOURINGMACHINES: An Menlo Park, Calif.: International Joint 578-583. Menlo Park. Calif.: International Joint Architecture for Dynamic, Rational, Mobile Conferences on Artificial Intelligence. Conferences on Artificial Intelligence. Agents, Ph.D. dissertation, Technical report Casteifranchi. C.; Micell, M.; and Cesra, A. Decker, K; Williamson. M.; and Sycara, K 273, Computer Laboratory, University of Cambridge. 1992. Dependence Relations among 1996a. Intelligent Adaptive Information Finin, T.; Fritzson, R.; McKay, D; and McEntire, Autonomous Agents. In Decentralized Artificial Agents. Paper presented at the AAAI-96 R. 1994. KQML as an Agent Cornmunication Intelligence, eds. S. Werner and Y. Demazeau, Workshop on Intelligent Adaptive Agents, Language. In Proceedings of the Third 215-231. Amsterdam, The Netherlands: 4-7 August, Portland. Oregon. International Conference on Information and Elsevier/North-Holland. Decker, K.; Williamson, M.; and Sycara, K Knowledge Management (CIKM94), 456-463. Chaib-draa, B. 1995. Industrial Applications of 1996b. Matchmaking and Brokering. In New York Association of Computing Distributed AI. Communications of the ACM Proceedings of the Second International Con- Machinery. 38(11): 49-53. ference on Multiagent Systems, 432. Menlo Park, Garrido, L., and Sycara, K 1996. Multiagent Chaib-draa, B.; Moulin, B.; Mandiau, R.; and Calif.: AAAI Press. Meeting Scheduling: Preliminary Millot, P. 1992. Trends in Distributed Dent, I.; Boticario, J.; McDermott, J.; Experimental Results. In Proceedings of the Artificial Intelligence. Artificial Intelligence Mitchell, T.; and Zabowski, D. 1992. A Per- Second International Conference on Multi- Review 6(1): 35-66. sonal Learning Apprentice. In Proceedings agent Systems, 95-102. Menlo Park, Calif.: AAAI Cohen, P. P., and Levesque, H. J. 1990. of the Tenth National Conference on Arti- Press. Intention Is Choice with Commitment. ficial Intelligence, 96-103. Menlo Park. Gasser, L. 1992. An Overview of DAI. In Artificial Intelligence 42(2-3): 213-261. Calif.: American Association for Artificial Distributed Artificial Intelligence Theory and Conry, S. L; Meyer, It. A.; and Lesser, V. R. Intelligence. Praxis, eds. N. Avouris and L. Gasser, 9-30. 1988. Multistage Negotiation in Distributed Drogul, A., and Ferber, J. 1992. From Tom Boston: Kluwer Academic. Planning. In Readings in Distributed Artificial Thumb to the Dockers: Some Experiments Gasser, L 1991. Social Conceptions of Knowledge and Action. Artificial Intelligence 47(1-3): 107- Intelligence, eds. A. H. Bond and L. Gasser, with Foraging Robots. In From Animals to 138. 367-384. San Francisco, Calif.: Morgan Animats: Second Conference on Simulation Gasser, L. 1986. The Integration of Kaufmann. of Adaptive Behavior, eds. H. K J. Meyer and Computing and Routine Work. ACM Corkill, D. D., and Lesser, V. It. 1983. The S. Wilson. Cambridge, Mass.: MIT Press. Transactions on Office Information Systems 4(3): Use of Metalevel Control for Coordination in Durfee. E. H. 1996. Planning in Distributed 205-225. a Distributed Problem- Solving Network. In Artificial Intelligence. In Foundations of Genesereth, M. R, and Ketchpel. S. P. 1994. Proceedings of the Eight International Joint Distributed Artificial Intelligence, eds. G. M. P. Software Agents. Communications of the ACM Conference on Artificial Intelligence (UCAI- O’Hare and N. K Jennings, 231-246. New 37(7): 48-53. 83), 767-770. Menlo Park. Calif.: York. Wiley. Grosz B, and Kraus, S.1996. Collaborative International Joint Conferences on Artificial Durfee, E. H. 1988. Coordination of Distributed Plans for Complex Group Actions. Artificial Intelligence. Problem Solvers. Boston: Kluwer Academic. Intelligence 86(2): 269-357. Cutkosky, M.; Fikes, L; Engelmore, R..; Durfee, E.H. 1987. A Unified Approach to Dynamic Coordination: Planning Actions and Grosz, B., and Sidner, C. 1990. Plans for Genesereth, M.; Mark, W.; Gruber, T.; Interactions in a Distributed Problem Solving Discourse. In Intentions in Communication, Tenenbaum, 3.; and Weber, J. 1993. PACT: An Network, Ph.D. dissertation, Department of eds. P. Cohen, 3. Morgan, and M. Pollack, Experiment in Integrating Concurrent Computer and Information Science, 417-444. Cambridge Mass.: MIT Press. Engineering Systems. IEEE Transactions on University of Massachusetts. Gruber, T. K 1993. A Translation Approach to Computers 26(1): 28-37. Durfee, E. H., and Lesser, V. 1989. Negotiat- Portable Ontologies. Knowledge Acquisition 5(2): Dan, T., and Birmingham, W. 1996. An ing Task Decomposition and Allocation 199-220. Attribute-Space Representation and Algo- Using Partial Global Planning. In Distributed Hardin, G. 1968. The Tragedy of Commons. rithm for Concurrent Engineering. Artificial Artificial Intelligence, Volume 2, eds. L Gasser Science 162(6): 1243-1248. Intelligence for Engineering Analysis, and Hewitt, C. 1986. Offices Are Open Systems. ACM Manufacturing 10(1): 21-35. and M. Huhns. 229-244. San Francisco, Calif.: Morgan Kaufmann. Transactions of Office Automation Systems Davis, R, and Smith, R. G. 1983. Negotiation 4(3): 271-287. as a Metaphor for Distributed Problem Dumfee, B. H.; Laser, V. R; and Corkill, D. D. Hu, J., and Wellman, M. P. 1996. Self-Fulfill- Solving. Artificial Intelligence 20(1): 63-100. 1987. Coherent Cooperation among ing Bias in Multiagent Learning. In Proceed- Decker, K, and Lesser, V. 1995. Designing a Communicating Problem Solvers. IEEE Family of Coordination Algorithms. In Pro- Transactions on Computers C-36(11): 1275-1291.

AI MAGAZINE

90 Twelfth International Workshop on Proceedings of the Second Pacific Rim Internat- ings of the Second International Conference on Distributed Artificial Intelligence (IWDAI- tional Conference on AI (PRICAI-92). 15-18 Multiagent Systems, 118-125. Menlo Park, Calif.: 93), 19-21 May, Hidden Valley, Pennsylva- September. Seoul, Korea. AAAI Press. nia. Loui, R. 1987. Defeat among Arguments: A Huberman, B, and Clearwater, S. H. 1995. A Kraus, S.; Wilkenfeld, J.; and Zlotkin. G. System of Defeasible Inference. Computational Multiagent System for Controlling Building 1995. Multiagent Negotiation under rime Intelligence 3(1): 100-106. Environments. In Proceedings of the First Constraints. Artificial Intelligence 75(2): 297-345. Maes, P. 1994. Agents That Reduce Work International Conference on Multiagent Systems, Lander, S.; Lesser. V. R.; and Connell, M. and Information Overload. Communications of 171-176. Menlo Park, Calif.: AAAI Press. E. 1991. Conflict-Resolution Strategies for the ACM 37(7): 30-40. Huberman, B A., and Hogg, T. 1988. The Maes, P. 1990. Designing Autonomous Agents. Behavior of Computational Ecologies. In The Cooperating Expert Agents. In CKBS-90, Cambridge. Mass.: MIT Press. Ecology of Computation. ed. B. A. Huberman. Proceedings of the International Working Con- Mason, C., and Johnson. K. 1989. DATMS: A Amsterdam, The Netherlands: North- ferarce on Cooperating Knowledge-Based Framework for Distributed Assumption-Based Holland. Systems, ed. S. M. Deen, 183-200. New York: Reasoning. In Distributed Artificial Intelligence Huhns, M. and Bridgeland, D. 1991. Multi-agent Springer-Verlag. Volume 2, eds. M. Hubris and L. Gasser, 293-318. Truth Maintenance. IEEE Transactions on Systems, Lesser, V. R. 1991. A Retrospective View of FA/C San Francisco, Calif.: Morgan Kaufmann. Man, and Cybernetics 216(6): 1437-1445. Distributed Problem Solving. IEEE Transactions Minsky, M. 1986. The Society of Mind. New Systems, Man, and Cybernetics. 21(6): 1347-1363. Huhns, M. and Singh, M. 1997. Readings in York Simon and Schuster. Lesser, V. K., and Corkill, D. D.1981. Functionally Agents. San Francisco, Calif.: Morgan Kaufmann. Accurate, Cooperative Distributed Systems. IEEE Mullen, T., and Wellman, M. P. 1996. Some Jennings, N. 1995. Controlling Cooperative Transactions on Systems, Man, and Cybernetics Issues in the Design of Market-Oriented Problem Solving in Industrial Multiagent Systems 11(1): 81-96. Agents. In Intelligent Agents II, eds. M. Using Joint Intention. Artificial Intelligence 75(2):

Wooldridge, J. P. Muller, and M© Tambe, 195-240. Lesser, V. R; Durfee, E. H. and Corkill, 283-298. Lecture Notes in Artificial Intelli- Jennings, N. ft.; Corera, J. M.; and Laresgoiti, D.1989. Trends in Cooperative Distributed gence 1037. New York Springer-Verlag. I. 1995. Developing Industrial Multiagent Systems. Problem Solving. IEEE Transactions on Muller, J.P., and Pischel, M. 1994. Modeling, In Proceedings of the First International Knowledge and Data Engineering 1(1): 63-83. Interacting Agents in Dynamic Environments. In Conference on Multiagent Systems, 423-430. Lewis. C. M., and Sycara. K 1993. Reaching Proceedings of the Eleventh European Conference Informed Agreement in Multispecialist Menlo Park, Calif.: AAAI Press. on Artificial Intelligence (ECAI-94) 709-713. Cooperation. Group Decision and Negotiation Jennings, N.; Sycara, K; and Wooldridge, M. Chichester, UK.: Wiley. 1998. A Roadmap for Agent Research and 2(3): 279-300. Myerson, R. B. 1989. Credible Negotiation Development. Autonomous Agents and Liu, 1.. and Sycara. K 1997. Coordination of Statements and Coherent Plans. Journal of Multiagent Systems. Multiple Agents for Production Management. Economic Theory 48 (2): 264-303. Jha, S.; Chalasani, Pr; Shehory, 0.; and Sycara, Annals of Operations Research 75:235-289. Newell, A. 1990. Unified Theories of Cognition, K 1998. A Formal Treatment of Distributed Liu, J., and Sycara. K 1995a. Emergent Boston: Harvard University Press. Matchmaking. In Proceedings of the Second Constraint Satisfaction through Multiagent O’Hare. G., and Jennings, N. 1996. Founda- International Conference on Autonomous Coordinated Interaction. In From Reaction to tions of Distributed Artificial Intelligence. New Agents (Agents ‘98). New York: Association Cognition, eds. C. Castlefranchi and J. P. York: Wiley. of Computing Machinery. Muller, 107-121. Lecture Notes in Artificial Parsons, S., and Jennings, N. R. 1996. Negotiation Jordan, J S. 1992. The Exponential Cover- Intelligence New York: Springer-Verlag. through Argumentation-A Preliminary Report. In gence of Bayesian Learning in Normal Form Liu, J., and Sycara, K P. 1995b. Exploiting Proceedings of the Second International Games. Games and Economic Behavior 4(2): Problem Structure for Distributed Coma- Conference on Multiagent Systems, 267-274. 202-217. Menlo Park. Calif.: AAAI Press. taint Optimization. In Proceedings of the Kephart, J. 0.; Hogg, T.; and Huberman, B. ¢¡¤£¦¥¨§ Parunak, V. 1987. Manufacturing Experience International Conference on Multiagent A. 1989. Dynamics of Computational with the Contract Net. In Distributed Artificial Systems, 246-253. Menlo Park, Calif.: AAAI Ecosystems: Implications for DAI. In Dis- Intelligence, Volume I, ed. M. Huhns, 285-310. Press. London: Pitman. tributed Artificial Intelligence, Volume 2, eds. L. Liu, J. and Sycara, K 1994. Distributed Rao, A. S., and George. M. P. 1991. Toward a Gasser and M. Huhns. 79-96. San Francisco, Problem Solving through Coordination in a Formal Theory of Deliberation and Its Role Calif.: Morgan Kaufmann. Society of Agents. In Proceedings of the in the Formation of Intentions. Technical Kinny, D.; Ljungberg, M.; Rao, A-; Sonen- Thirteenth International Workshop on Distributed Report, Australian Artificial Intelligence berg. L; Tidhard, G.; and Werner, E. 1992. Artificial Intelligence, 190-206. Menlo Park, Institute, Victoria, Australia. Planned Team Activity. In Artificial Social Calif.: AAAI Press. Rick, C.. and Sidner, C. 1997. COLLAGEN: Systems, eds. C. Castelfranchi and E. Werner. Liu, J., and Sycara, K P. 1993. Collective When Agents Collaborate with People. In New York: Springer-Verlag. Problem Solving through Coordinated Proceedings of the First International Kornfeld, W. A., and Hewitt, C. E. 1981. The Reaction. In Proceedings of the IEEE Inter- Conference on Autonomous Agents (Agents Scientific Community Metaphor. IEEE national Conference on Evolutionary ‘97), 284-291. New York: Association of Transactions on Systems, Man, and Cybernetics Computation. 575-578. Washington. D.C.: Computing Machinery. 11(1): 24-33. Institute of Electrical and Electronics Engi- Rosenschein, J. S., and Genesereth, M. R. Kraus, S.; Nirke. M.; and Sycara. K. 1993. neers. 1985. Deals among Rational Agents. In Reaching Agreements trough Argumentation: Ljunberg. M., and Lucas, A. 1992. The A Logical Model. Paper presented at the Oasis Traffic Management System. In

SUMMER 1998

91 Proceedings of the Ninth International Joint Stone, P., and Veloso, M. 1997. Multiagent Adversarial Environment. In Proceedings of Conference on Artificial Intelligence (IJCAI- Systems: A Survey from the Machine-Learning the First International Workshop on 85), 91-99. Menlo Park, Calif.: International Perspective, Technical report, CMU-CS-97-193, ROBOCUP. San Francisco, Calif.: Morgan Joint Conferences on Artificial Intelligence. School of Computer Science, Carnegie Mellon Kaufmann. University. Rosenschein, J. S., and Zlotkin, G. 1994. Rules Wang, H., and Wang, C. 1997. Intelligent Sycara, K. 1997. Using Option Pricing to Value of Encounter: Designing Conventions for Agents in the Nuclear Industry. IEEE Automated Negotiation among Computers. Commitment Flexibility in Multiagent Systems, Computer 30(11): 28-34. Technical report, CMU-CS-TR-97-169, School of Cambridge, Mass.: MIT Press. Weihmayer. R., and Velthuijsen, H. 1994. Sandholm, T. 1993. An Implementation of the Computer Science, Carnegie Mellon University. Application of Distributed AI and Cooperative Contract Net Protocol Based on Marginal Sycara, K. 1991. Problem Restructuring in Problem Solving to Telecommunications. In Cost Calculations. In Proceedings of the Negotiation. Management Science 37(10): 1248- AI Approaches to Telecommunications and Eleventh National Conference on Artificial 1268. Intelligence. 2S6-262. Menlo Park, Calif.: Sycara, K. 1990a. Negotiation Planning: An AI Network Management, eds. J. Liebowitz and D. American Association for Artificial Approach. European Journal of Operational Prereau. Amsterdam, The Netherlands: 105 Intelligence. Research 46(2): 216-234. Press. Sandholm, T., and Lesser, V. 1995. Issues in Sycara. K. 1990b. Persuasive Argumentation in Williamson, M.; Decker, K.; and Sycara. K. Negotiation. Theory. and Decisions 2&203-242. Automated Negotiation and Electronic Commerce: 1996. Unified Information and Control in Extending the Contract Net Protocol. In Sycara. K. P. 1988. Resolving Goal Conflicts via Hierarchical Task Networks, in Proceedings Negotiation. In Proceedings of the Seventh Proceedings of the Second International Conference of the AAAI-96 Workshop on Theories of National Conference on Artificial Intelligence on Multiagent Systems, 328-335. Menlo Park, Planning, Action, and Control. Menlo Park, (AAAI-88). Menlo Park, Calif.: American Calif.: AAAI Press. Calif: AAAI Press. Schwuttke, U. M., and Quan, A. G. 1993. Association for Artificial Intelligence. Sycara. K. 1987. Resolving Adversarial Conflicts: Wooldridge. M., and Jennings, N. 1995. Enhancing Performance of Cooperating Agents in An Approach to Integrating Case Based and Intelligent Agents: Theory and Practice. Real-Time Diagnostic Systems. In Proceedings of Analytic Methods. Ph.D. dissertation, School of Knowledge Engineering Review 10(2): 115-152. the Thirteenth International Joint Conference on Information and Computer Science, Georgia Zeng, D., and Sycara, K. 1998. Bayesian Learning Artificial Intelligence (IJCAI-93), 332-337. Menlo Institute of Technology. in Negotiation. International Journal of Human- Park, Calif.: International Joint Conferences on Sycara. K.; Decker, K.; and Zeng, D. 1998. Computer Studies 48:125-141. Zeng, D., and Artificial Intelligence. Intelligent Agents in Portfolio Management. In Sycara, K. 1997. Benefits of Learning in Sen, S., ed. 1996. Working Notes of the 1996 Agent Technology Foundations, Negotiation. In Proceedings of the National AAAI Symposium on Adaptation, Coevolution Applications, and Markets. eds. M. Conference on Artificial Intelligence (AAAI- and Learning in Multiagent Systems. Menlo Wooldridge and N. R. Jennings, 267-283. 97), 36-41. Menlo Park, Calif.: American Park, Calif.: American Association for Artificial Berlin: Springer. Association for Artificial Intelligence. Intelligence. Sycara, K.; Roth, S.; Sadeh, N.; and Fox, M. 1991. Zlotkin, G., and Rosenschein, J. 1991. Distributed Constrained Heuristic Search. Shehory, 0.; Jha. S.; and Sycara, K. 1997. Cooperation and Conflict Resolution via IEEE Transactions on System, Man, and Multiagent Coordination through Coalition Negotiation among Autonomous Agents In Cybernetics 21(6): 1446-1461. IEEE Transactions Formation. In Agent IV Agent Theories, Noncooperative Domains. Sycara, K.; Decker, K.; Pannu, A.; on Systems, Man, and Cybernetics Architectures, and Languages. eds. M. Singh, A. (Special Issue Williamson, M.; and Zeng, D. 1996. Dis- Rao, and M. Wooldridge, 143-154. Lecture on Distributed Artificial Intelligence) 21(6): tributed Intelligent Agents. IEEE Expert Notes in Artificial Intelligence. New York 1317-1324. 11(6): 36-46. Springer-Verlag. Tambe, M. 1997. Toward flexible Team- Shoham, Y. 1993. Agent-Oriented Program- Katia P. Sycara is an associate research work. Journal of Artificial Intelligence Research ming. Artificial Intelligence 60(1): 51-92. professor In the School of Computer Science 7(1): 83-124. Simon, H. 1957. Models of Man: Social and at Carnegie Mellon University. She is also Rational-Mathematical Essays on Rational Human Thomas, J., and Sycara, K. 1998. Stability the director of the Advanced Behavior in a Social Setting. New York Wiley. and Heterogeneity In Multiagent Systems. In Information Technology Laboratory. She is Proceedings of the Third International Con- Smith, I.. and Cohen, P. 1996. Toward directing-conducting research and ference on Multiagent Systems. Menlo Park, Semantics for an Agent Communication developing technology in multiagent systems Calif: AAAI Press. Language Based on Speech Acts. In Pro- to support the decision making of multiple Travers, M. 1988. Animal Construction Kit ceedings of the Fourteenth National Con- decision makers. For the past four years, she In Artificial Life, ed. C. G. Langton, 421-442. ference on Artificial Intelligence (AAAI-96). has led research in the development of Reading, Mass.: Addison-Wesley. Van Menlo Park, Calif.: American Association for distributed collaborating intelligent agents Dyke. Artificial Intelligence. that interact with their users, other agents, Parunak, H. 1987. Manufacturing Steeb, R., Cammarata, S.; Hayes-Roth, F.; and distributed information resources to Experience with the Contract Net. In Thorndyke. P.; and Wesson, R. 1988. Dis- assist their users in information finding and Distributed Artificial Intelligence, ed. M. tributed Intelligence for Air Fleet ControL In 51-tering and the planning and execution of Huhns, 285-310. London: Pitman. Readings in Distributed Artificial Intelli- various tasks. gence, eds. A. H. Bond and L. Gasser, 90- Veloso, M.; Stone, P.; Han, K.; and Achim, 101. San Francisco, Calif.: Morgan S. 1997. CMUNITED: A Team of Robotic Kaufmann. Soccer Agents Collaborating in an

AI MAGAZINE