Mediators in the Architecture of Future Information Systems

Gio Wiederhold

Stanford University

Septemb er

An edited version of this rep ort was published in

The IEEE Computer Magazine March

Abstract

The installation of highsp eed networks using optical b er and high bandwidth messsage

forwarding gateways is changing the physical capabilities of information systems These

capabilities must b e complemented with corresp onding software systems advances to obtain

a real b enet Without smart software we will gain access to more data but not improve

access to the typ e and quality of information needed for decision making

To develop the concepts needed for future information systems we mo del information

pro cessing as an interaction of data and knowledge This mo del provides criteria for a

highlevel functional partitioning These partitions are mapp ed into information pro cess

ing mo dules The mo dules are assigned to no des of the distributed information systems A

central role is assigned to mo dules that mediate between the users workstations and data re

sources Mediators contain the administrative and technical knowledge to create information

needed for decisionmaking Software which mediates is common to day but the structure the

interfaces and implementations vary greatly so that automation of integration is awkward

By formalizing and implementing mediation we establish a partitioned information sys

tems architecture which is of managable complexity and can deliver much of the p ower

that technology puts into our reach The partitions and mo dules map into the p owerful

distributed hardw are that is b ecoming available We refer to the mo dules that p erform these

services in a sharable and comp osable wayasmediators

We will present conceptual requirements that must b e placed on mediators to assure

eective largescale information systems The mo dularity in this architecture is not only

a goal but also enables the goal to b e reached since these systems will need autonomous

mo dules to p ermit growth and enable them to survive in a rapidly changing world

The intent of this pap er is to provide a conceptual framework for many distinct eorts

The concepts provide a direction for an information pro cessing systems in the foreseeable

future We also indicate some subtasks that are of research concern to us In the long range

the exp erience gathered by diverse eorts may lead to a new layer of highlevel communication

standards

Intro duction

Computerbased information systems connected to worldwide highsp eed networks provide

increasingly rapid access to a wide variety of data resources Mayo This technology op ens

up p ossibilities of access to data requiring capabilities for assimilation and analysis which

greatly exceed what wenowhave in hand Without intelligent pro cessing these advances

will have only a minor b enet to the user at a decisionmaking level That brave user will

be swamp ed with illdened data of unknown origin

The problems

We nd twotyp es of problems for single databases a primary hindrance for enduser access

is the volume of data that is b ecoming available the lack of abstraction and the need to

understand the representation of data the ma jor concern when combining information from

multiple databases the mismatch encountered in information representation and structure

We will expand on those two issues

Volume and Abstraction

The volume of data can b e reduced by selection It is not coincidental that SELECT is the

principal op eration of relational database management systems but selected data is still at

to o ne a level of detail to b e useful for decision making Further reduction is achieved by

bringing data to higher levels of abstraction Aggregation op erations as COUNT AVERAGE

SD MAX MIN etc provide some computational faciliti es for abstraction but anysuch

abstraction is formulated within the application using some domain knowledge

Typ e of abstraction Example

Base data Abstraction

Granularity Sales detail Product summaries

Generalization Product data Product type

Temp oral Daily sales Seasonally adjusted monthly sales

Relative Product cost Inflation adjusted trends

Exception recognition Accounting detail Evidence of fraud

Path computation Airline schedules Trip duration and cost

Figure Abstraction functions

For most base data more than one abstraction must b e supp orted for the salesmanager

the aggregation is by sales region while for marketing aggregations by customer income are

appropriate Examples of required abstraction typ es are given in Fig

Computational requirements for abstraction are often complicated The groupings to

dene abstractions may for instance involve recursive closures These cannot b e sp ecied

with current database query languages Application programs are then written byspe

cialists to reduce the data The use of sp ecic datapro cessing programs as intermediaries

diminishes exibility and resp onsiveness for the end user Now the knowledge that creates

the abstractions is hidden and hard to share and reuse

Mismatch

Data obtained from remote and autonomous sources will often not match in terms of naming

scop e granularity of abstractions temp oral bases and domain denitions as listed in Figure

The dierences shown in the examples must b e resolved b efore automatic pro cessing can

join these values

Typ e of mismatch Example

Domains

Key dierence Alan TuringThe Enigma versus QATH

reference for reader reference for librarian

Scop e dierence employees paid versus employees available

includes retirees includes consultants

Abstraction grain personal income versus family income

from employment for taxation

Temp oral basis monthly budget versus weekly production

central oce factory

Domain semantics postal codes versus town names

one can cover multiple places can have multiple codes

Value semantics excessive pay versus excessive pay

per internal revenue servive per board of directors

Figure Mismatches in data resources

Without an extended pro cessing paradigm as prop osed in this pap er the information

needed to initiate actions will b e hidden in ever larger volumes of detail scrollable on ever

larger screens in ever smaller fonts In essence the gap b etween information and data will

yet b e wider than it is now Knowing that information exists and is accessible creates

exp ectations by endusers Finding that it is not available in a useful form or that it cannot

b e combined with other data creates confusion and frustration Webelieve that the ob jection

made by some users ab out computerbased systems that they create information overload

is voiced b ecause those users get to o much of the wrong kind of data

Use of a model

In order to visualize the requirements we place on future information systems we consider

the activitie s that are carried out to day whenever decisions are b eing made The making of

informed decisions requires the application of a variety of knowledge to information ab out

the state of the world To clarify the distinction of data and and knowledge in our mo del we

restate a denition from WiederholdB

Data describ es sp ecic instances and events Data may gathered automatically or clerically

The correctness of data can b e veried visavis the real world

Knowledge describ es abstract classes Each class typically covers many instances Exp erts

are needed to gather and formalize knowledge Data can b e used to disproveknowledge

To manage the variety of knowledge we employ sp ecialists and to manage the volume of

data we segment our databases In these partitions partial results are pro duced abstracted

and ltered The problem of making decisions is now reduced to the issue of chosing and

evaluating the signicance of the pieces of information derived in those partitions and fusing

the imp ortant p ortions

For example an investment decision for a manufacturer will dep end on the fusion of

information on its own pro duction capabilityversus that of others of its sales exp erience in

related pro ducts of the market for the conceived pro duct at a certain price of the costto

price ratio appropriate for the size of the market and its cost of the funds to b e invested

Sp ecialists will consider these diverse topics and each sp ecialist will consult multiple data

resources to supp ort their claims The decision maker will fuse that information using

considerations of risk and longrange ob jectives in combining the results

An eective architecture for future information systems must supp ort automated in

formation acquisition pro cesses for decisionmaking activiti es By default we approachthe

solution in an manner analogous seen in humanbased supp ort systems While we mo del the

system based on abstractions from the world of human information pro cessing wedonot

constrain the architecture to b e anthrop omorphic and to mimic human b ehavior There are

many asp ects of human b ehavior that wehave not yet b een able to capture and formalize

adequately Our goal is merely to dene an architecture wholly comp osed of pieces of soft

ware that are available or app ear to b e attainable in a mo dest timeframe saytenyears The

demands of the software in terms of pro cessing cost should b e such that mo dern hardware

can deal with it

Current state

We are not starting from a zero base Systems are b ecoming available now with the capabil

ities envisaged byVannevar Bush for his memex in Bush We can select and scroll

information on our workstation displays wehave remote access to data and can presentthe

values on one of multiple windows we can insert do cuments into les in our workstation

and we can annotate text and graphics We can reach conclusions based on this evidence

and advise others of decisions made

Our vision in this pap er is intended to carry us b eyond sp ecically to provide a basis

for automated integration of such information Our current systems as well as memexdo

not adress that phase

A Mo del of Information Pro cessing

The ob jective of obtaining information was already clearly stated byWoodward Wood

ward Information enables us to decide among several actions which are otherwise not

distinguisable Consider again a simple business environment A factory manager needs sales

data to set pro duction levels A sales manager needs demographic information to pro ject

future sales A customer wants price and quality information to make purchase choices

Most of the information needed by the p eople in the example can b e represented by fac

orks tual data and should b e available on some computer somewhere Communication netw

have the p otential to make data available wherever it is needed However to make the deci

sions a considerable amountofknowledge has to b e applied as well Today most knowledge

is made available through a variety of administrative and technical sta in an institution

Waldrop Some knowledge is enco ded in datapro cessing programs and exp ert systems

for automated pro cessing

The pro cess of generating supp orting information do es not dier much in partially au

tomated or manual systems In manual systems the decision maker obtains assistance from

sta and colleagues who p eruse les and prepare summarizations and do cumentation for their

advice In automated systems the sta is likely to use computers to prepare these do cuments

The decisionmaker rarely uses the computer b ecause the information from multiple sources

is dicult to integrate automatically

The pro cessing and the application of knowledge

Atechnician will knowhow to select and transfer data from a remote computer to one

used for analysis A data analyst will understand the attributes of the data and dene

the functions to combine and integrate the data deMichiel A statistician mayprovide

pro cedures to aggregate data on customers into groups that present distinctivebehavior

patterns A psychologist mayprovide classication parameters that characterize the groups

Finally a manager has to assess the validity of the classications that have b een made

use the information to make a decision and assume the risk of making the decision A

public relations p erson may take the information and present it in a manner that can b e

explained to the sto ckholders to whom the risk is eventually distributed Since these tasks

are characterized by using data and knowledge gathered in the past with the ob jectiveof

aecting the future we will refer to these tasks as planning Our denition of planning is

broader than that used in Articial Intelligence research Cohen although the ob jectives

are the same To b e able to deal with these planning actions in a fo cused waywe will mo del

the information pro cessing asp ects

There is a recurring activity here

Data are made available These are either factual observations or results from prior

pro cessing and combinations thereof

Knowledge is made available It derives from formal training and exp erience

Knowledge ab out the data and its use is applied in two phases

i Selection subsets of available data are

dened obtained and merged

ii Reduction the data found are summarized to an appropriate level of abstraction

Several such results are made available and fused

oways The combined information is utilized in tw

i Actions are taken that will aect the state of the world

ii Unexp ected results will b e used to augment the exp erience base of the participants

and others who receive this information increasing their knowledge

The pro cess lo ops are closed when

i The actions and their eect is observed and recorded in some database

ii The knowledge is recorded to aect subsequent data denition selection or fusion

The two distinct feedback lo ops and their interaction are illustrated in Figure The data

lo op closes when the eects of actions taken are recorded in the database The knowledge

lo op closes when recently gained knowledge is amde available so it can b e used for further

selection and data reduction decisions

The interaction is of prime concern for future systems To ols for the other steps are easy

to identifywe list some in Section Wenow consider in detail asp ects of Step identied

ab ove Data and Knowledge

Knowledge Loop Data Loop Information is created at the Storage Education confluence of data -- the state Selection Recording & Integration knowledge -- the ability to Abstraction select and Experience State changes project the Decision-making state into Action the future

Exp erience Data acquisition

Education Data validation

Knowledge Stored Data

selection

knowledge data

Information

feedback feedback

lo op lo op

reduction by

summarization

abstraction

integration

ranking and

pruning

Decision Supp ort

induction actions

learning state changes

Figure The knowledge and data feedback lo ops and their interaction

Creation of information

Let us identify where during pro cessing information is created Since getting information

means that something has b een obtained that was not known prior to the event one or more

of the following conditions have to hold

The information is obtained from a remote source and was previously not known

lo cally Step i ab ove Here the information system must provide communication

supp ort

A sp ecial case of this condition is when a database is used for recall to pro

vide data we knew once but cannot rememb er with certainty Here the database

comp onent is used to communicate over time from the past to the present

Two facts previously distinct are merged or unied Step i ab ove A classic

although trivial example is nding ones grandfather via transitivity of parents In

realistic systems unication of data also involves computation of functions saythe

average income and its variation of groups of consumers

Multiple results are fused using pragmatic assessments of the quality and risks

asso ciated within Step Here abstractions are combined rather than facts and the

pro cessing techniques are those asso ciated with symb olic pro cessing in rulebased

exp ert systems although they are also found co ded in application programs In

our example the market sp ecialist maywant to unify incomes of current consumers

with their reading habits to devise an advertising strategy

Databases record detailed data for eachofmany instances or events Reducing this detail to

a few abstract cases raises the information content p er element Of these abstractions only

a small feasible number of justied results is brought to the decisionmaker For instance

the market analyst has made it p ossible that decisions do not have to deal with individual

consumers but with consumer groups Certain groups may b e unlikely purchasers and are

not targeted for promotions

While the b ehavior of any individual may not b e according to the rules hyp othesized

in the prior steps the exp ected b ehavior of the aggregate p opulation should b e close to the

prediction Failures o ccur only if wehave many errors in the underlying data or serious

errors in our knowledge Uncertaintyhowever is common

Uncertainty

We cannot predict the future with certaintyFor automation of fullscale information systems

the pro cessing of uncertaintymust b e supp orted although there are subtasks whichcan

b e precisely dened Uncertainties within a domain may b e captured by a formal mo del

ty computation can b e subsumed by pro Although wehave the argument that all uncertain

babilistic reasoning Cheeseman Horvitz it seems that the variety of assumptions

made is based on dierences in domain semantics During analysis uncertain events or states

have to b e combined for extrap olation into the future We still havenooverall mo del to

predict which uncertaintycombining computations will b e b est for some domain

To assess the extent of uncertainty aecting predictions wemust combine the uncertain

precision of source information and the uncertainty created at each step where wecombine

them In the various steps we collect observations based on some criterion say p eople

living in a certain p ostalco de area Wealsohave data to asso ciate an income level with that

p ostalco de area and p erhaps even an income distribution At the same time wemayknow

the income distribution of p eople buying some pro duct

It is hence natural to make the conceptual unication step of p ostal co de to p otential

sales Unfortunately there is no logical reason why such a unication should b e correct

Wehave some formal classes namely p eople with a certain p ostal co de and some other

formalizable classes based on income in addition there are some natural classes not formal

ized but intuitively known In our example some natural classes of interest are the p otential

purchasers of which there are several subgroups namely those that b ought in the past those

that will buy to day and those that will b e buying the planned pro ducts For the future class

only informal criteria can b e formulated

The planner at the decisionmaking no de will use denable classes by p ostal co de

by observed and recorded purchasing patterns to establish candidate memb ers for the

natural class of p otential consumers These classes overlap the b etter the overlap the

more correct the decisionmakers predictions will b e If we infer over classes that do not

matchwell the uncertainty attached to the generated plans will b e greater But uncertainty

is the essence of decisionmaking and is reected in the risk that the manager of our example

takes on Wewould not need a manager with decisionmaking skills if weonlyhave to rep ort

the p ostal co des for our base group of p otential consumers

Summary of the information mo del

Eective information is created at the conuence of knowledge and data For prediction we

use knowledge to conceive rules applicable to natural classes Selection of data for decision

making is constrained by b eing based on formal class denitions Uncertainty is created

when formal and natural classes are matched

Communication of knowledge and data is necessary to achieve this conuence The

er time The information systems we consider communication may o ccur over space or ov

must supp ort b oth communication and fusion of data and knowledge

Our systems must also b e able to deal with continuing change Both data and knowl

edge change over time b ecause the world changes and b ecause we learn things ab out our

world Rules that were valid once eventually b ecome riddled with exceptions and a sp ecial

ist who do es not adapt will nd his work to b ecome without value An information system

architecture must deal explicitl y with knowledge maintenance

Information System Comp onents

We will nowcharacterize the comp onents available to day to build information systems All

these comp onents are p ositioned along a data highwayprovided by mo dern communication

technology The interactions of the comp onents will b e primarily constrained by logical and

informational limitations and not byphysical linkages When we place the comp onents into

the conceptual architecture we will recognize lacunae ie places where there are currently no

existing only inadequate or unco op erative comp onents We will see where the comp onents

must work together Eective systems can b e achieved only if the data and knowledge

interfaces of the comp onents are such that co op eration is feasible

Data and knowledge resources

There is a wide variety of data resources Wemight classify them by a measure of closeness

to the source Raw data obtained from sensors such as purchase records obtained by p oint

ofsale scanners or on a dierent scale images recorded by earth satellites are at the factual

extreme Census and sto ck rep orts contain data that have seen some pro cessing but will b e

considered as facts by most users

At the other extreme are textual representations of knowledge Bo oks research rep orts

and library material in general contain knowledge contributed by the writers and their col

legues Unfortunately from our pro cessingoriented view that knowledge is not exploitable

ts are without a human mediator The text tables and gures contained in such do cumen

data as well but rarely in a form that can b e transcrib ed for database pro cessing

If do cument information is stored in bibliographic systems suchas dialog Summit

or medline Sewell only selection and presentation op erations are enabled Textual

material do es not lend itself well to automated analysis abstraction and generalization

Some typ es of rep orts pro duced routinelytendtohave a degree of structure that makes

some extent of analysis feasible as demonstrated for chemical data Callahan pathology

rep orts Sager or military event rep orting messages McCune

Workstation applications

The systems environment for planning activiti es is provided by the new generations of ca

pable workstations For planning the users need to interact with their own hyp otheses save

intermediate results for comparison of eects and display the alternate pro jections over time

This interaction with information establishes the insights needed to gain condence in ones

decisions

The user exercises creativity at the workstation We hence do not try to constrain the

user here at all By providing comprehensive supp ort for access to information we hop e that

the complexity of the endusers applications can b e reduced to suchanextent that quite

involved analyses remain manageable

Mo dern op erating and network systems help much with simplifying the users tasks by

removing concern ab out managing hardware resources The architecture presented here is

meant to address the management of information resources residing on such hardware

The base information for planning pro cesses has to b e obtained from a variety of data

resources A capable interface is required

Mediation

An interface from the users workstation to the database servers which only denes com

munication proto cols and formats in terms of database elements do es not deal with the

abstraction and representation problems existing in to days data and knowledge resources

The interfaces must take on an active role

We will refer to the dynamic interface function as mediation This term includes the

pro cessing needed to maketheinterfaces work the knowledge structures that drivethe

transformations needed to transform data to information and anyintermediate storage that

is needed

To clarify the term we will list some examples of mediation found in current information

systems This list could b e greatly expanded in length and depth The citations provide

linkages for followup we exp ect that most readers will have encountered these or similar

functions Typ es of mediation functions that have b een develop ed are

Transformation and subsetting of databases using view denitions and ob ject tem

plates Chamb erlin Wiederhold Lai Barsalou

Metho ds to access and merge data from multiple databases Smith Dayal

Sacca

Computations that supp ort abstraction and generalization over underlying data

Hammer Adiba Ozsoyoglu Downs Wiederhold deZegher

Cichetti Chen Chaudhuri

Intelligent directories to information bases such as library catalogs indexing aids

and thesaurus structures Humphrey Doszkos McCarthy

Metho ds to deal with uncertainty and missing data b ecause of incomplete or mis

matched sources Callahan Chiang Litwin deMichiel

Many more examples can b e added Most readers will have used or created suchinterface

mo dules at some time since we all need information from large autonomous and mismatched

sources A ma jor motivation for exp ert database systems is mediation We excluded from the

list simple pro cessing techniques that do not dep end on knowledge that is extraneous to the

database prop er such as indexes caching mechanisms network services and le directories

The examples of mediation shown are sp ecialized and tied either to a sp ecic database

or to a particular application or b oth We will now dene mediators as mo dules o ccupying

an explicit activelayer b etween the users applications and the data resources Our goal is

a sharable architecture Domain-specific Mediation

• User application – Workstations • Mediator – Expert-owned nodes • Data sources – Remote primary and byproduct services

The Mediator Architecture

In order to provide intelli gent and active mediation weenvisage a class of software mo dules

which mediate b etween the workstation applications and the databases These mediators

will form a distinct middle layer making the user applications indep endent of the data

resources What are the transforms needed in suchalayer and what form will the mo dules

supp orting this layer have The resp onses to these questions are interrelated We will rst

identify problems with generalizing the concept of mediation identied in current information

systems and then dene some general criteria We will also justify the necessityofhaving a

mo dular domainsp ecic organization in this layer

The Architectural Layers

We create a central layer by distinguishing the function of mediation from the useroriented

pro cessing and from database access Most user tasks will need multiple distinct mediators

for their subtasks A mediator uses one or a few databases

The interfaces that are to b e supp orted provide the cuts where communication network

services are needed as shown in Figure

result decision making

Layer Indep endent applications on workstations managed by decision makers

network services to information servers

Layer Multiple mediators managed by domain sp ecialists

network services to data servers

Layer Multiple databases managed by database administrators

input  realworld changes

Figure The three layers of this architecture

Unfortunately the commonality of function seen in the examples cited in Section

do es not extend to an architectural commonality the various examples are b ound to the data

resources and the endusers applications in dierentways It is here where new technology

must b e established if fusion at the application level is to b e supp orted Accessing one

mediator at a time do es not allow for fusion and seeing multiple results on distinct windows

of one screen do es not supp ort automation of fusion

An inappropriate approach to mediation

The concept of mediation is related to the notion of having corp orate information centers

as promoted by IBM and others Atre These are to b e corp orate resources staed and

equipp ed with to ols to aid any users needing information The needs and capabilities of an

information center however dier in two resp ects from those of automated mediators

A single center or mediator for that matter cannot deal with all the varieties of

information that is useful for corp orate decisionmaking

Automation of the function will b e necessary to achieve acceptable resp onse time

and growth of knowledge and its qualityover time

The user should not b e burdened with the task of seeking out information sources

This task esp ecially if it is rep etetive is b est left to an interaction of application

programs in a workstation and automated mediation programs

The information center notion initiates yet another selfserving bureaucracy within a cor

p oration Eective sta is not likely to b e contentintheinternal service roles that an

information center provides so that turnover of sta and its knowledge will b e high The

only role foreseen in such a center is mediation bringing together p otential users with

candidate information

In order to manage mediation mo dularity instead of centralization seems to b e essential

Mo dularity is naturally supp orted by a distributed environment the dominantenvironment

of computing in the near future

Mediators

Now that wehave listed some examples of mediators in use or planned for sp ecic tasks we

can give a general denition

A mediator is a software mo dule that exploits enco ded knowledge ab out some sets or subsets

of data to create information for a higher layer of applications

We place the same requirements on a mediation mo dule that we place on any software

mo dule it should b e small and simple Bo ehm Bull so that it can b e maintained by

one exp ert or at most by a coherent group of exp erts

An imp ortant although p erhaps not essential requirementwed like to place on medi

ators is that they b e insp ectable by the p otential users For instance the rules used bya

mediator using exp ert system technology can b e obtained by the user as in any good coop

erative exp ert system Wick In this sense the mediators provide data ab out themselves

in resp onse to insp ection and such data could b e analyzed byyet another an insp ector

mediator

Since eventually there will b e a great numb er and variety of mediators the users have

to b e able to cho ose among them Insp ectability enables that task For instance distinct

mediators which can provide the ten b est consultants for database design may use dierent

evaluation criteria one may use the numb er of publications and another the number of

clients

Some metamediators will have to exist that merely provide access to catalogues listing

available mediators and data resources The searchmay go either way for a given data

source one maywant to lo cate a knowledgeable mediator and for a desirable mediator we

need to lo cate an adequate data resource It will b e essential that the faciliti es provided by

these metalevel mediators can b e integrated into the general pro cessing mo del since search

for information is always an imp ortant asp ect of information pro cessing Where search

and analyses are separated as is still common to day for instance in statistical data

pro cessing trying to nd the data is often the most costly phase of information pro cessing

For databases that are autonomous it is desirable that only a limited and recognizable

set of mediators dep end on anyone of them Fo cusing data access through a limited number

of views maintained by these mediators provides the data indep endence which is necessary for

databases that are evolving autonomously Currently compatibility constraints are hindering

growth of databases in terms of structure and scop e since many users are aected As the

numb er of users and the automation of access increases the imp ortance of indirect access

via mediators will increase

The Interface to Mediators

The most critical asp ect of this threelayer architecture are the twointerfaces that are now

created Todays mediating programs employ a wide varietyofinterface metho ds and ap

proaches The user learns to use one or a few of them and then remains committed until

its p erformance b ecomes wholly unacceptable Unless the mediators are easily and exibly

accessible the mo del of common information access weenvisage is b ound to remain a ction

It is then in the in terface and its supp ort that the researchchallenge lies Since our hardware

environment implies that mediators can liveonany no des not only on the workstations and

database hosts their interfaces must b e grounded in communication proto cols

The Users Workstation Interface to the Mediators

The range of capabilities that mediators mayhaveissuch that a highlevel language should

evolve to drive the mediators We are thinking of language concepts here rather than

of interface standards to indicate the degree of extensibili ty that must b e provided if the

mediating concepts are to b e generalized

Determining an eectiveinterface b etween the workstation application and the mediators

will b e a ma jor research eort in rening this mo del It app ears that a language is needed

to provide exibility comp osability iteration and evaluation in this interface Descriptive

but static interface sp ecications do not seem to b e able to deal with the variety of control

and information ow that must b e supp orted The basic language structure should p ermit

incremental growth so that new functions can b e supp orted as mediators join the network

to provide new functionality

It is imp ortant to observe that wedonotseeaneedforauserfriendly interface We

need here a machine and communicationfriendly interface Application programs execut

ing on the users workstations can provide the typ e of display and manipulation functions

appropriate to their typ es of user Omitting the criterium of userfriendliness avoids the

dichotomy that has lead to inadequacies in sql which tries to b e userfriendly while its

predominant use will b e for programmed access Stonebraker Standards needed here can

only b e dened after exp erience has b een obtained in sharing of these resources to supp ort

the highlevel functions needed for decisionmaking

The Mediator to DBMS Interface

Existing database standards as sql and rda provide a basis for database access by mediators

Relational concepts such as selection views etc provide an adequate starting p oint A

mediator dealing mainly with a sp ecic database need not b e constrained to a particular

interface proto col while a mediator which is more general will b e eective through a standard

interface A mediator whichcombines information from multiple databases may use its

knowledge to control the merging pro cess and use a relational algebra subset Joins may

for instance b e replaced by explicit semijoins so that intelligent ltering can o ccur during

pro cessing Still dealing with multiple sources is likely to lead to incompleteness Outer

joins are often required for data access to avoid losing ob jects with incomplete information

Wiederhold

The separation of user applications and databases that the mediating mo dules provide

ws reorganization of data structures and redistribution of data over the no des without also allo

aecting the functionality of the mo dules The threelevel architecture then makes an explicit

tradeo in favor of exibilityversus integration The arguments for this tradeo fo cus on

the variety of uses made of databases and by extension the results of mediator mo dules

Wiederhold

Sharability of information requires that database results can b e congured according

to one of several views Mediators b eing active can create ob jects for a wide variety

of orthagonal views Barsalou

On the other hand making complex ob jects themselves p ersistent binds knowledge

to the data which hinders sharability Maier

The loss of p erformance due to the interp osition of a mediator can b e overcome

bytechniques listed in Section

These arguments do not yet address the distribution of the mediators weenvisage

Available interfaces

Weneedinterface proto cols for data and knowledge For data transmission there are devel

oping standards The level of concern for mediation is within the application layer in the

OSI sense Tanenbaum Wehave faith that communication systems will so on handle all

lower level supp ort layers without ma jor problems The Remote Data Access RDA proto col

provides one such instance ISORDA

Other technologies that are pushing capabilities at the interface is the National File

System b eing promoted by Sp ector at CMU However as mentioned in the intro duction

communication of data alone do es not guarantee that the data will b e correctly understo o d

for pro cessing by the receiver Dierences often exist in the meaning assigned to the bits

stored some examples were shown in Figure

Sharing of mediator mo dules

In either case since we are getting access to so much more data from a variety of sources

arriving at ever higher rates automated pro cessing will b e essential The pro cessing tasks

needed within the mediators are those sketched in the interaction mo del of Fig selection

fusion reduction abstraction and generalization Diverse mediator mo dules will use these

functions in varying extents to provide the supp ort for user applications at the decision

making layer ab ove

The mediator mo dules will b e most eective if they can servea variety of applications

Hay esRoth The applications will comp ose their tasks as much as p ossible by acquiring

information from the set of available mediators Unavailable information may motivate the

creation of new mediators

Sharing reinforces the need for twotyp es of partitioning one into horizontal layers

for endusers mediators and databases resp ectively and two vertically into multiple user

applications each using various congurations of mediators A mediator in turn will use

distinct views over one or several databases Just as databases are justied by the shared

usage they receive mediators should b e sharable Note that to days exp ert systems are

rarely mo dular and sharable so that their development and maintenance cost is harder to

amortize

For instance the mediation mo dule which can deal with ination adjustmentcanbe

used by many applications The mediator which understands p ostal co des and town names

can b e used by the p ost oce express delivery services and corp orate mail ro oms

We foresee here an incentiveforavariety of sp ecialists to develop p owerful but generally

useful mediators which can b e used bymultiple customers Placing ones knowledge into a

mediator can b e more rapidly eective and p erhaps more rewarding than writing a b o ok

on the topic

We can now summarize these observations in Figure

User User User User User Useri Userk Usern

Query Relevant resp onses Insp ection

Mediatorj Mediatork Mediatorl Mediatorm

Formatted query Bulky resp onses Triggered events

Database w Database x Database y Database z

All are distributed over nationwide networks

modules

Figure Interfaces for information ow

Distribution of Mediators

Wehave implied throughout that mediators are distinct mo dules distributed over the net

work Distribution can b e motivated by greater economy for access by b etter lo cality for

maintenance and by issues of mo dularityFor mediators the two latter arguments are the

driving force for distribution

Why should mediators not b e attached to the databases In many cases it maybe

feasible in general it is not appropriate

The mediator contains knowledge that is b eyond the scop e of the database prop er

A database programmer dealing with say a factory pro duction control system

cannot b e exp ected to foresee all the strategic uses of the collected information

Concepts of abstraction are not part of database technology to day the fo cus has

b een on reliable and consistent management of large volumes of detailed facts

Intelligent pro cessing of data will often involve dealing with uncertaintyadding

excessive complexity to database technology

Many mediators will access multiple databases to combine disjointed sets of data

prior to analysis and reduction

Similarily we can argue that the mediators should not b e attached to the users worksta

tion applications Again the functions that mediators provide are of a dierent scop e than

the tasks b eing p erformed on the workstations Workstation applications mayusea variety

of mediators to explore the data resources

A ma jor motivation for keeping mediators distinct is maintenance During the initial

stage of most pro jects whichdevelop ed exp ert systems their knowledge bases simply grew

and the cost of knowledge acquisition dominated the cost of knowledge maintenance Many

pro jects in fact assumed implicitly that knowledge once obtained would remain valid for

all times History teaches us that this is not true although some fundamental rules may

indeed not change for a long time the application heuristics for its use change as the demands

of the world change Concepts as knossTsichritzis recognize the problem and fo cus on

the problem within a sp ecic domain

Maintenance by an outside exp ert of knowledge stored within an application system is

intrusive and risky The enduser should make the decision when new knowledge should b e

incorp orated We hence keep the mediator mo dules distinct

The eciency concerns of separating knowledge and data can b e mitigated by replication

Since mediators incorp orating only knowledge and no factual data are relatively stable

they can b e replicated as needed and placed on no des along the data highway where they

are maximally eective They certainly should not change during a transaction As long as

the mediators remain small they can also b e easily shipp ed to a site where substantial data

volumes have to b e pro cessed

Triggers for knowledge maintenance

Wehave added one asp ect of mediation into Figure which has not yet b een discussed Since

the knowledge in the mediator must b e kept uptodate it will b e wise for many mediators to

place triggers or active demons into the databases Buneman Stonebraker Nowthe

mediators can b e informed when the database and by extension the realworld changes

The owner of the mediator should ensure that suchchanges are in time reected in the

mediators knowledge base

We consider again justied by the analogy to human sp ecialists that a mediator is

fundamentally trusted but is insp ectable when suspicions of obsoleteness arise For instance

an assumption say that welltodo p eople buy big cars may b e used in the marketing

mediator but it is p ossible that over time this rule b ecomes invalid We exp ect then that the

base data b e monitored for changes and that exceptions to database constraints will trigger

information ow to the mediator In a rulebased mediator the certainty factor of some rule

can b e adjusted Esculier If the uncertainty exceeds a threshold the mediator can advise

its creator the domain exp ert to abandon this rule The enduser need not get involved

Risch

Related Research

Wehave shown throughout that many ofthe individual concepts underlying this architecture

have b een found in earler work Thiis is hardly suprising since the problems that future

information systems must address exist now and are b eing dealt with in many sp ecic

situations Our contribution is a generalization of the practice and a fo cusing of more general

research to these systems Wedowant to p oint out some signicant relationships as well as

our own fo cus

Indep endent actors and agents

There is an obvious corollary b etween the mediators prop osed here and the concept of actors

Hewitt However a considerable constraint is imp osed on our mediators namely they

do not interact intelligently with each other so that a hierarchy is imp osed for every sp ecic

task The reason for this constraint is to provide computational simplicity and manageability

An actor mo del can of course b e used within a mediator to implement its computational

task

The network of connections within the global architecture means that distinct tasks can

intersect at no des within this information pro cessing structure The same mediator typ e may

access distinct sets of data and information from one data source can b e used by distinct

mediators

Hierarchical task control

Automation requires an understanding of the control mechanisms that maintain balance

and motivate progress or growth To what extentnetworks of autonomous agents can b e

motivated is unclear

Rather than extrap olating into the unknown we dene an architecture that w ecan

conceptually manage to day and keep our minds op en to extensions that are b eyond to days

conceptual foundations We take again a cue here from Vannevar Bush who could identify

all units needed for the memex although its comp onents were based on technology that did

not exist in

Todays organizations dep end greatly on hierarchical structures Many information pro

cessing functions now carried out in organizations are p erformed bylower levels to obtain

aggregate use and rank information for the decisionmaking levels of management Current

development of the actors mo del Hewitt stresses the necessary matchbetween actors and

functions seen in realworld organizations The org mo del also follows an anthrop omorphic

paradigm Malone fo cusing on a wider set of problemsolving interactions and provides

insights into distributed mediation In all this research concepts that mo del understo o d

organizational practices form a basis a notion that is broadly argued by Litwin

The relationships of users to mediators is organizationally similar to that presented in

the contract net mo del by Smith although the fo cus of a mediator is on partitioning for

management and not on leastcost of computations Contract net concepts mayprovide ideas

for charging and accounting in this environment an issue not discussed in this pap er but

dealt with in the fast Cohen pro ject The distributed co op erating agents of Ko o

deal with nonadversary contracting a very desirable mo del for a future world

Maintenance and learning

In eective organizations lower levels of managementinvolved in informationpro cessing also

provide feedback to sup erior layers

The knowledge emb o died in the mediators cannot b e assumed to b e static Some knowl

edge may b e up dateable byhuman exp erts but for activeknowledge domains some au

tomation will b e imp ortant Providing advice on inconsistencies b etween acquired data and

assumed knowledge is the rst step

Eventually some mediators will b e endowed with learning mechanisms Feedback for

learning may either come from p erformance measures or from explicit induction over the data

bases they manage Blum Wilkins The trigger for learning is monitoring Changes

in the database can continuously up date hyp otheses of interest within the mediator The

validity of these hyp otheses can b e assessed by insp ecting corollary observations in the view

of the mediator DeZegher The uncertaintyofhyp otheses can provide a ranking when

an application task requests assessments

Techniques

Mediators will embody a variety tec hniques now found b oth in freestanding applications and

in programs that p erform mediation functions These programs are now often classied by

the underlying scientic area rather than by its place in information systems

Techniques from articial intelli gence

The nature of mediators is such that many techniques develop ed in AI will b e employed We

exp ect that mediators will often use

Declarative approaches

Capability for explanation

Heuristic control of inference

Pruning of candidate solutions

Evaluating the certainty of results

The literature on these topics is broad Cohen Heuristic approaches are likely to b e

imp ortant b ecause of the large solution spaces Uncertainty computations are needed to deal

with missing data and mismatched classes

Techniques from logical databases

Since the use of mediators encourages a formal approach to the pro cessing of data techniques

b eing develop ed within logical and deductive database are likely to b e equally imp ortant As

long as conventional DBMS do not provide well for recursive search the computations that

achieve closure are likely to b e placed into mediators Beeri Ramakrishnan Since

prop er denition of the rules for stable and nite recursion is likely to remain dicult these

techniques are b est managed by exp erts

Another example of logical mediation is dealing with generalization in order to return

results of sp ecied cardinality Chaudhuri A frequent abstraction is to derive temp o

ral interval representations from detailed event data Here we also see a need to attach

techniques that dep end on domain semantics to the attributes in the database Ja jo dia

Techniques for ecient access

In this pap er we do not fo cus on the eciency of mediated access We realize that interp osing

alayer in our information systems is likely to have a signicant cost We will argue that

the exibility and adaptability of a mo dular approachwillovercome in time the ineciencies

induced by an inexible rigid structure The partitioning of the tasks will also make research

subtasks more managaable It is to day infeasible to carry out a really signicantintegrated

system developmentinany single academic institution Within sp ecialized lab oratories sub

stantial systems can b e develop ed and tested but those are still hard to integrate into the

world outside

ercoming pro cessing We can show some examples of systems research that fo cuses on ov

b ottlenecks

Caching and materialized views and view indexes within the mediators Roussop ou

los Hanson Sheth

Rule bases for semantic query optimization King Chakravarthy

data reorganization to follow dominant access demands Fursin

an ability to abandon ineective ob ject bindings Giord

Privacy protection for sensitive data through interface mo dules Cohen

Sharing of these Techniques

Articial intellige nce logical and systems techniques can of course b e shared among the

mediators Only knowledge sp ecic to the application needs to b e lo cal to each mediator

In the framework whichwe presenta varietyoftechniques can co op erate and comp ete to

improve the pro duction of information

SoDs

The KBMS Pro ject group at Stanford is in the pro cess of formulating a sp ecic form of

mediators called SoDs Wiederhold A SoD provides a declarative structure for the

domain semantics suitable for an interpreter We see multiple SoDs b eing used byan

application executing a long and interactive transaction Wesummarize our concepts here

as one researchavenue in providing comp onents for the architecture wehave presented

Sp ecic features and constraints imp osed on SoDs are

The knowledge should b e represented in a declarative form

SoDs should haveawellformed language interface

They contain feature descriptions exploitable by the interpreter

They should b e insp ectable by the user applications

They should b e amenable to parallel execution

They access the databases through relational views

During execution source and derived data ob jects are b ound internally

They consider ob ject sharing

By placing these constraints on SoDs as mediators we hop e to b e able to prove asp ects of

their b ehavior and interaction Provable b ehavior is not only of interest to develop ers but

also provides a basis for prediction of computational eciency

However the mo dularity of SoDs causes twotyp es of losses

Loss in p ower due to limitations in interconnections

Loss in p erformance due to reliance on symb olic binding rather than on direct

linkages

ving structures that enable We hop e to oset these losses through gains obtained byha

eective computational algorithms An implementation of a demonstration using frame

technology is b eing expanded The longrange b enet is of course that small wellconstructed

mediators will enable knowledge maintenance and growth

An Interface Language

Our researchidenties the language problem as a ma jor issue If we cannot express the high

level concepts encapsulated in the mediators well then we will not b e able to implement

the required servbices For application access to SoDs we start from database concepts

where highlevel languages have b ecome accepted WWHC The SoD access language

SoDaL will have the functional capabilities of sql plus iteration and test to provide Turing

machine level capability Qian New predicates are needed to sp ecify intelligent selection

Selection of desirable ob jects requires an ability to rank ob jects according to sp ecied criteria

These criteria are understo o d by the SoD and are not necessarily predicates on underlying

data elements although for a trivial SoD that may b e true These criteria are asso ciated with

result size parameters as Give me the b est X where the b est predicate is a semantically

overloaded term interpreted internally to a particular SoD

The format of SoDaL is not envisaged to b e userfriendly after all other subsystems

will use the SoDs not p eople It should have a clear and symmetric structure whichis

machine friendly It should b e easy to build a userfriendly interface if needed on top of a

capable SoDaL

Limits and Extensions

The separation into layers reduces the exibility of information transfer Esp ecially struc

turing the mediators into a single layer b etween application and data is overly simplistic

Precursors to general mediators already recognize hierarchies as the contract net Smith

and general interaction as actors Hewitt

A desire to serve largescale applications is the reason for the simple architecture pre

sented here To assure eective exploitati on of the mediator concepts we prop ose to intro duce

complexity within their layer and the asso ciated pro cessing cost slowly onlyasthefoun

dations are established to p ermit ecientuse

Structuring mediators into hierarchies should not lead to problems We already assumed

that directory mediators could insp ect other mediators Insp ection of lowerlevel mediators

is also straightforward Lowlevel mediators may only have database access knowledge and

understand little application domain semantics On the other hand highlevel mediators can

take on minor decisionmaking functions

More complex is lateral information sharing among mediators Some such sharing will b e

needed to maintain the lexicons that preserve ob ject identity when distinct mediators group

and classify data Optimizers may restructure the information ow taking into account

success or failure with certain ob jects in one of the involved SoDs

Fully general interaction among mediators is not likely to b e supp orted at this level of

abstraction Just as human organizations are willing to structure and constrain interactions

even at some lostopp ortunity cost we imp ose similar constraints on the broad information

systems weenvisage

Learning by mo difying certainty parameters in the knowledgebase is relatively simple

ehavenomechanisms that relate Learning of new concepts is much more dicult since w

observations automatically to unsp ecied symb olic concepts By dep ending initiall y fully on

the human exp ert to maintain the mediator then moving to some parameterization of rule

priorities we can gradually move to automated learning

Eciency is always a concern Once derived data are available the need to analyze large

databases is reduced but the intermediate knowledge has to b e maintained Binding of the

knowledge to provide the eect of compilation of queries is an imp ortant strategy Work

in rulebased optimizers and automatic creation of exp ert systems p oints in that direction

Scho en These tactics exacerbate the problems of maintaining integrity under concurrent

use Researchinto truthmaintenance is relevant here Filman Kanth

Requirements of data securitymay imp ose further constraints Dealing with trusted

mediators however may encourage database owners to participate in information sharing to

a greater extent than they would if all participants would need to b e granted lelevel access

privileges

Summary

Weenvisage a variety of information pro cessing mo dules residing in no des along the data

highways that advances in communication technology can nowprovide A conceptual lay

ering distinguishes no des by function decisionmaking exploration information supp ort by

mediation of data by knowledge and base data resources

The mediation function now seen in a variety of programs is placed into explicit me

diation mo dules or mediators For claritywe place all the mediators into one horizontal

layer These mediators are to b e limited in scop e and size to enable maintenance byan

exp ert as well as insp ection and selection by the enduser Mediators are asso ciated with the

domain exp ert but may b e replicated and shipp ed to other network no des to increase their

eectiveness Sp ecialization increases the p ower and maintainability of the mediators and

provides choices for the users

Applications obtain information by dealing with abstractions supp orted by the media

tors and not by accessing base data directly A language will b e needed to provide exibility

in the interaction b etween the endusers workstation and the mediators We discuss the par

titioning of articial intelli gence paradigms into pragmatics at the userworkstation layer

and the formal infrastructure in the mediation layer further in Wiederhold

For query op erations the control ow go es from the application to the mediator who

would interpret the query to plan optimal database access The data would ow to the medi

ator b e aggregated reduced pruned etc and the results rep orted to the query originator

Multiple mediators serve an application with pieces of information from their sub domains

tinintelligent mediators indicate the critical The knowledgebased paradigms inheren

role of articial intelli gence technology foreseen when implementing mediators Knowledge

sources of the mediation examples listed in Section range from typ e information to business

rules The mediating mo dules we are developing SoDs stress structure and and declarative

domain semantics for interpretation and insp ection

Mediators may b e strengthened byhaving learning capability Derived information may

simply b e stored in a mediator Learning can also lead to new tactics of data acquisition and

control of pro cessing

The intent of the architectural mo del is not to b e exclusive and rigid It is intended to

provide a common framework under which many new technologies can b e accommo dated

As shown throughout many existing concepts can b e viewed as implementations of media

tion Esp ecially the techniques listed in Section havevalidity within and outside of this

framework but will b ecome more accessible and sharable Now they are at b est enb edded

within intelligent application programs

An imp ortant ob jective of the architecture is the ability to utilize a variety of infor

mation sources without demanding that they b e broughtinto a common format and with

only minimal requirements on their interfaces Maintenance of the knowledge bases in the

mediators requires sp ecializati on and a manageable size

In a recent rep ort the three primary issues to b e addressed in knowledgebased

systems were maintenance problem mo deling and learning and knowledge acquisition

Buchanan The architecture wepresented here contributes to all three issues largely

byproviding a partitioning that p ermits large systems to b e comp osed from mo dules that

are maintainable that can implement sp ecic submo dels and that access domain data for

learning and knowledge validation

Acknowledgement

This pap er integrates and extends concepts develop ed during researchinto the managementof

large knowledge bases primarily supp orted byDARPA under contract NC Useful

insights were gathered byinteraction with researchers at DEC pro ject title Reasoning ab out

rme and with the national HIVmo deling community Cohen The DARPA Principal

Investigators meeting Nov Dallas TX help ed with solidifying these concepts in part by

providing a mo dern view of the communication and pro cessing capabilities that lie in our

future Rob ert Kahn of the Corp for National Research Initiatives encouraged development

of these ideas Further inputs were provided by panel participants at the DASFAA Seoul

April VLDB Amsterdam August and IFIP San Francisco August

conferences Researchwork on triggers is b eing carried out byTore Rischatthe

HewlettPackard Stanford Science Center and his input help ed clarify salient p oints Andreas

Paepke of HP commented as well Dr Witold Litwin of INRIA and the students on the

Stanford KBMS pro ject esp ecially Sura jit Chaudhuri provided a critical review and helpful

comments

References

Adiba Michel E Adiba Derived Relations A Unied Mechanism for Views Snap

shots and Distributed Data VLDB Zaniolo and Delob eleds Sep pp

Atre Shaku Atre Information Center Strategies and Case StudiesVol Atre Int

Consultants Rye NY

Barsalou Thierry Barsalou An Ob jectbased Arc hitecture for Biomedical Exp ert

Database Systems SCAMC IEEE CS Press Washington DC Novemb er

Basu Amit Basu Knowledge Views in Multiuser Knowledge Based Systems IEEE

Data Engineering Conference Feb Los Angeles

Beeri C Beeri and R Ramakishnan On the Power of Magic ACMPODS San

Diego Mar

Blum Rob ert L Blum Discovery and Representation of Causal Relationships from a

Large TimeOriented Clinical Database The RX Pro ject Springer Verlag Lecture

Notes in Medical Informatics no

Bo ehm Barry W Bo ehm Software Engineering Economics IEEE Trans Software

EngVol No Jan pp

Bull M Bull R Duda D Port and J Reiter Applying Software Engineering

Principles to KnowledgeBase Development Pro c Exp ert Systems and Business

NY Learned Information Meadford NJ Nov pp

Buchanan BG Buchanan D Bobrow R Davis J McDermott and EH Short

lie Research Directions in KnowledgeBased Systems Stanford KSL rep ort

to app ear in J Traub ed Annual Review of Computer Science

Buneman OP Buneman and EK Clemons Eciently Monitoring Relational

Databases ACM Trans on Database SystemsVol No Sep pp

Bush Vannevar Bush As WeMay Think Atlantic MonthlyVol No

pp

Callahan MV Callahan and PF Rusch Online implementation of the CA

SEARCH le and the CAS Registry Nomenclature File Online RevVol No

Oct pp

Chamb erlin DD Chamb erlin JN Gray and IL Traiger Views Authorization

and Lo cking in a Relational Data Base System Pro c NCC AFIPS Vol

AFIPS Press pp

Chakravarthy US Chakravarthy D Fishmann and J Minker Semantic Query Op

timization in Exp ert Systems and Database Systems Exp ert Databases Kersch

b erged Benjamin Cummins

Chaudhuri S Chaudhuri Generalization and a Framework for Query Mo dica

tion IEEE Data Engineering Los Angeles Feb

Cheeseman Peter Cheeseman In Defense of Probability Pro c IJCAI Los Ange

les pp

Chen MC Chen and L McNamee A Data Mo del and Access Metho d for Sum

mary Data Management IEEE Data Engineering Conf Los Angeles Feb

Chiang TC Chiang and GR Rose Design and Implementation of a Pro duction

hnical JournalVol Database Management System DBM Bell System Tec

No Nov pp

Cicchetti R Cicchetti LD Lakhal N LeThanh and S Miranda A Logical

Summarydata Mo del for Macro Statistical Databases DASFAA Seoul Korea

KISS and IPSJ Apr pp

Cohen PR Cohen and E Feigenbaum eds The Handb o ok of Articial Intelli

gence Morgan Kaufman

Cohen H Cohen and S Layne editors Future Data Management and Access

Workshop to Develop Recommendations for the National Scientic Eort on AIDS

Mo deling and Epidemiology sp onsored by the White House Domestic Policy Coun

cil

Dayal U Dayal and HY Hwang View Denition and Generalization for Database

Integration in Multibase A System for Heterogeneous Databases IEEE Trans

Software EngVolSE No Nov pp

DeMichiel Linda DeMichiel Performing Op erations over Mismatched Domains

IEEE Data Engineering Conference Feb IEEE Transactions on Knowledge

and Data EngineeringVol No Dec

DeZegher I DeZegherGeets AG Freeman MG Walker RL Blum and G Wie

derhold Summarization and Display of Online Medical Records MD Comput

ingVol no March pp

Downs SM Downs MG Walker and RL Blum Automated summarization of

online medical records IFIP Medinfo NorthHolland pp

Doszko cs Tamas E Doszko cs Natural Language Pro cessing in Information Re

trieval JAmSo cInfSciVol No Jul pp

Esculier Christian Esculier Intro duction a la Tolerance Semantique PhD thesis

Un Joseph Fournier Grenoble

Filman Rob ert E Filman Reasoning with Worlds and Truth Maintenance in

a KnowledgeBased Programming Environments Comm ACMVol No

Apr pp

Fursin Gennadiy I Fursin Estimating and Decision Making Techniques in Data

base Design Control Systems and Machines Kiev No Jan pp

Giord DK Giord RM Needham and MD Schro eder The cedar File Sys

tem Comm ACMVol no Mar pp

Gray Jim N Gray An Approach to Decentralized Computer Systems IEEE

Trans Software EngVolSe No Jun pp

Hammer M Hammer and D McLeo d The Semantic Data Mo del A Mo delling

Machanism for Data Base Applications Pro c ACM SIGMOD Lowenthal and

Daleeds pp

Hanson Eric Hanson A Performance Analysis of View Materializati on Strategies

Pro c ACMSIGMOD May

HayesRoth FrederickHayesRoth The Knowledgebased Exp ert System A tuto

rial IEEE Computer Sep pp

Hewitt C Hewitt P Bishop and R Steiger A Universal Mo dular actor Formal

ism for Articial Intelligence IJCAI SRI Aug pp

Hewitt Carl Hewitt Knowledge Pro cessing Organizations prop osal Nov

Horvitz EJ Horvitz DE Heckerman and C Langlotz A Framework for Compar

ing Alternate Formalisms for Plausible Reasoning Pro c AAAI pp

Humphrey SM Humphrey A Kap o or D Mendez and M Dorsey The Indexing

Aid Pro ject Knowledgebased Indexing of the Medical Literature NLM LH

NCBC Mar

ISORDA Working draft of ISO Remote Access Proto col ISOTCSC N

ANSI XH Jul

Kanth MR Kanth and PK Bose Extending an Assumptionbased Truth Main

tenance System to Databases IEEE CS Data Engineering Conference Feb

LosAngeles

Kaplan SJerrold Kaplan Designing a Portable Natural Language Database Query

System ACM TODS Vol No Mar pp

King Jonathan J King Query Optimization by Semantic Reasoning Univof Michi

gan Press

Ko o CC Ko o and G Wiederhold A Commitmentbased Communication Mo del

for Distributed Oce Environments ACMCOIS Mar pp

Lai KY Lai TW Malone and KC Yu Ob ject Lens A Spreadsheet for Co op er

ativeWork ACM Ttans on Oce Inf SystemsVol No Oct pp

Litwin W Litwin and A Ab dellatif Multidatabase Interop erability IEEE Com

puterVol No Dec pp

Litwin W Litwin and N Roussop olous A Mo del for Computer Life University

of Maryland Institute for Advanced Computer Studies UMIACSTR July

Maier David Maier Why isnt there an Ob jectoriented Data Mo del Information

Pro cessing Ritter ed IFIPS NorthHolland pp

Malone TW Malone KR Grant FA Turbak SA Brobst and MD Cohen

Intelligent InformationSharing Systems Comm ACMVol No May

pp

Mayo JS Mayo and WB Marx jr Intro duction T echnology of Future Net

works ATT Technical JournalVol No Mar

McCarthy John L McCarthy Knowledge Engineering or Engineering Information

Do We Need New To ols IEEE Data Engineering Conference Feb Los

Angeles

McCune BP McCune RM Tong JS Dean and DG Shapiro rubric A

System for Rulebased Information Retrieval IEEE Trans Software EngVolSE

no Sep pp

McIntyre SC McIntyre and LF Higgins Knowledge Base Partitioning For Lo cal

Exp ertise Exp erience In A Knowledge Based Marketing DSS Hawaii Conf on

Inf Systems Feb

Ozsoyoglu ZM Ozsoyoglu and G Ozsoyoglu SummaryTableByExample A

Database Query Language for Manipulating Summary Data IEEE Data Engi

neering Conf Los Angeles Apr

Qian XiaoLei Qian The DeductiveSynthesis of IterativeTransaction PhD thesis

Stanford University June

Ramakrishnan Raghu Ramakrishnan Conlog Logic Control Un Wisconsin

Madison CSD

Risch Tore Risch Monitoring Database Ob jects Pro c VLDB Amsterdam

Aug Morgan Kaufmann Pubs

Roussop oulos N Roussop oulos and H Kang Principles and Techniques in the

Design of ADMS IEEE ComputerVol No Dec pp

Sacca D Sacca D Vermeir A dAtri A Liso SG Pedersen JJ Snijders and

N Spyratos Description of the Overall Architecture of the kiwi System ES

PRIT EEC Elseviers pp

Sager N Sager EC Chi C Friedman and MS Lyman Mo delling Natural

Language Data for Automatic Creation of a Database from FreeText Input IEEE

Database Engineering BullVolNo Sep pp

Scho en E Scho en RG Smith and BG Buchanan Design of Knowledgebased

Vol Systems with a Knowledgebased Assistant IEEE Trans Software Eng

No Dec pp

Sewell W Sewell and S Tietelbaum Observations of Enduser Online Searching

Behavior over Eleven Years JAmSo cInfSci Vol No Jul pp

Shen Sheldon Shen Design of a Virtual Database Information Systems Vol

No pp Rj Query pro cessing involves building a query tree with

Sheth AP Sheth JA Larson and A Cornellio A To ol for Integrating Conceptual

Schemas and User Views IEEE Data Engineering Conference Feb Los

Angeles

Smith RG Smith The Contract Net Proto col HighLevel Communication and

Control in a Distributed Problem Solver IEEE Ttans ComputersVolC No

Dec pp

Smith JM Smith et al multibase Integrating Heterogeneous Distributed

Database Systems Pro cNCCAFIPSVol Mar pp

Stonebraker M Stonebraker D DuBourdieux and W Edwards Triggers and In

ference in Database Systems Bro die Mylop oulos and Schmidt eds On Knowl

edge Base Management Systems Integrating Articial Intelligence and Database

Technologies Springer Feb pp

Stonebraker Michael Stonebraker Future Trends In Database Systems IEEE

Data Engineering Conf Feb Los Angeles

Summit RK Summit dialog An Op erational OnLine Reference Retrieval Sys

tem ACM Nat Conf pp

Tanenbaum Andrew S Tanenbaum Computer Networks nd ed PrenticeHall

Tsichritzis D Tsichritzis E Fiume S Gibbs and O Nierstrasz knos Knowledge

Acquisition Dissemination and Manipulation Ob jects ACM Trans on Oce Inf

SysVol No Jan pp

Waldrop MMitchell Waldrop The Intelligence of Organizations ScienceVol

No Sep pp

Wetherb e JC Wetherb e and RL Leitheiser Information Centers A Survey of

Services Decisions Problems and Successes Inf Sys ManagVol No

pp

Wick MR Wick and JR Slagle An Explanation Facility for Todays Exp ert

Systems IEEE Exp ert Spring pp

Wiederhold Gio Wiederhold Database Design McGra wHill

WiederholdB Gio Wiederhold Knowledge versus Data Chapter of Bro die

Mylop oulos and Schmidt eds On Knowledge Base Management Systems In

tegrating Articial Intelligence and Database Technologies Springer Verlag June

pages

Wiederhold Gio Wiederhold Views Ob jects and Databases IEEE Computer

Vol No Decemb er Pages

Wiederhold G Wiederhold and XL Qian Mo deling Asynchrony in Distributed

Databases IEEE Data Engineering Conference Los Angeles Feb

Wiederhold G Wiederhold P Rathmann T Barsalou BS Lee and D Quass

Partitioning and Combining Knowledge Information Systems to app ear

Woodward Woodward Philip M Probability and Information Theory with Ap

plications to RadarMcGrawhill Pergamon

WWBD GCM Wiederhold MG Walker RL Blum and SM Downs Acqui

sition of Knowledge from Data Pro c ACM SIGART ISMIS Oct pp

WWHC GCM Wiederhold MG Walker W Hasan S Chaudhuri A Swami

SK Cha XL Qian M Winslett L DeMichiel and PK Rathmann KSYS

An Architecture for Integrating Databases and Knowledge Bases in Amar Gupta

ed Heterogenous Integrated Information Systems IEEE Press

Wilkins DC Wilkins WJ Clancey and BG Buchanan Knowledge Base Re

nementby Monitoring Abstract Control Knowledge Stanford TR STANCS

Aug