Download From: AAAI Technical Report S-9 -0 . Compilation copyright © 199

From: AAAI Technical Report WS-93-07. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved.
Exploring the role of single function agents in negotiation
D. C. Brown
AI Research Group
ComputerScience Department
WorcesterPolytechnic Institute
Worcester MA01609
USA
BRIEF POSITION PAPER
Despite somevery interesting workon computationalmodelsof negotiation such as that by
Sycara [1990], Lander & Lesser [1992], Klein [1991], Werkman
[1992], Kannapan& Marshek
[1992], and others we feel that there is still muchto be done. Despite having contacted
manypeople directly and via email, the first surprise is that there appear to be very few
implementations, few computational models, little evaluation of the systems, and no
comparativeevaluation.
As our group is newto the area, we are first doing a survey of workon Negotiation. Weare
attempting to characterize most of the mainresearch efforts by developingand using a set of
attributes. Theseinclude the presence and role of a central agent; the amountof knowledge
each agent has about other agents; the communication
languageused; whetherthere is a global
record of the design; the reasons for conflict; whetherthe agents are cooperative; whetherthe
agents are homogeneous;whether a design/negotiation history is kept/used; and the
negotiationstrategies used.
Ourmotivation for the research is the desire to build workingConcurrentEngineering(CE)
systems. Withinthe last year wehavebuilt two systemswith manyCEcharacteristics. As it is
difficult to producea precise description of CEon whichall researchers wouldagree, I will
refrain from claiming that they are CEsystems. However,most wouldagree that the essential
quality is that at designtime it is importantto considerthe impactof the designon aspects from
later in the life-cycle, suchas packaging,or recycling.
Oneof the systemsbuilt, the 13Dsystem[Bauschet al 1992][Victor et al 1993], used multiple
cooperative intelligent agents to serve on a design team with a humandesigner. Togetherthey
designed simple powderceramic components,using a conceptual design phase and a detailed
design phase. Theagents were able to respondand assist in both the design phases, offering
cost estimation, material selection, process simulation and inspection planning, for example.
Therewere16 intelligent agents in I3D.
47
The second system, called SNEAKERS
[Douglas & Brown1993], was intended to be used
during training courses to help educate users about the powerof CEand its importance. As
such, it is a single-user system, with the other membersof the CEteam being simulated by
expert systems (agents). The user designs a tower on the screen, according to some
requirements,by selecting from, and specializing, parts offered by the system. As in I3D, the
agents offer a variety of comments
whichthe user can act on or ignore.
Thesesystems have several things in common.Thoserelevant to this workshopare: they used
intelligent agents, with each agent havinga single function; the agents are only allowedto
activate in a fixed sequence.
In these systemsall possibility of conflict, and hencenegotiation, wascompiledout by careful
ordering of the agents and by careful groupingof the decisions being made.In a newversion
of 13D(13D+),to address grouping, weare separating out the decisions into different agents,
and, to addressordering, weare trying not to predefinetheir order of execution.
The work in I3D+is currently in progress. It will provide a testbed for a variety of
experiments with negotiation. Wewill also be constructing a customizable domain
independenttestbed to with whichto experiment.
In 13Dand SNEAKERS
we experimented with single function agents. Agents were only able
to perform one of a few simple roles [Brown1992]. The roles used included providing
Advice,Analysis, Criticism, Estimation, Evaluation, Planning, Selection and Suggestion.
These roles can be contrasted with the target about which they can makecomments.For I3D
these include Material, Process, Manufacturing,Inspection, Cost, Reliability and Durability.
Theroles and targets define a 2Dmatrix of possible agents, not all of whichare sensible.
In I3D+we are trying to refine this further by considering another axis, point-of-view, thus
classifying agents using a 3Dmatrix (i.e., role, target, point-of-view).For example,one agent
might be Criticism of material (the Targe0 from the Point-of-view of cost. Note that, as
before, not every entry is sensible. However,this systemallows us to quite precisely define
everyagent in the system.It will also force us to be very explicit aboutwhateach agent knows,
while allowing us freedomto implementthe agents in any wayappropriate, providedthey play
their appointedrole.
Weintend to extend this sort of analysis to apply to the knowledgeused in negotiation. Each
role in the negotiation process will be played by a separate agent, with, for example,some
beingresponsiblefor providingtrade-off knowledge,or keepingthe negotiation history. In this
waywe hope to develop a detailed architecture that supports negotiation, but which will
responddifferently whenprovided with different knowledge.
Weexpect to learn a lot aboutthe roles of different types of knowledge
in different negotiation
schemes. Where grouping of decisions allowed us to avoid negotiation in I3D and
SNEAKERS,
this approachof strict separation will force us to face negotiation directly and in
detail. It is alreadyclear that Critics andEstimatorsinteract in interesting wayswith the agents
doing design (Advisors), allowing several waysto negotiate betweenthem. For example,
critic can criticize an estimate: what do they negotiate about? Perhaps the methodof
48
estimation? Or the data used?
In this first round of investigation into negotiation we will study the architecture that supports
negotiation, rather than negotiation strategies. Wewill mainly focus on using I3D as the
testbed, modifyingand extending it as necessary.
In the longer term, we are very interested in the use of design histories during negotiation, in
case-based negotiation, in the use of trade-off knowledge, in comparative analyses of
negotiation strategies, and in the opportunities for use of Learning during negotiation.
References
J.J.Bausch, D.C.Zenger, D.C.Brown, R.Ludwig & R.D.Sisson (August 1992) Integrated
Design and Manufacturing Strategies for Powder~ Processing Applications. Proceedings,
University
Programs in Computer Aided Engineering
Design and Manufacturing
(UPCAEDM’92),
Tennessee Technological University.
D.C.Brown(1992) Design. Encyclopedia of Artificial
(Ed.), J.Wiley.
Intelligence,
2nd edn., S.C.Shapiro
R.E.Douglas, Jr. & D.C.Brown(1993) A Concurrent Engineering Demonstration System for
use in Training Engineers and Managers. Proc. AI in Engineering Conference.
S.M.Kannapan & K.M.Marshek (July 1992) A Schema for Negotiation between Intelligent
Design Agents in Concurrent Engineering. In: D.C.Brown, M.Waldron & H.Yoshikawa
(Eds.), Intelligent ComputerAided Design, Elsevier Science Publishers B.V. (North-Holland).
M.Klein (1991) Supporting conflict resolution in cooperative design systems. IEEE Trans.
Systems, Man, and Cybernetics, November/December,Vol. 21, No. 6, pp. 1379-1390.
S.E.Lander & V.R.Lesser (Dec. 1992) Customizing Distributed Search among Agents with
Heterogeneous Knowledge. Proc. 5th Int. Symp. on AI Applications in Manuf. & Robotics,
Cancun, Mexico.
K.P.Sycara (1990) Cooperative Negotiation
Engineering Design. Springer Verlag.
in Concurrent
Design.
In: Cooperative
S.K.Victor, D.C.Brown, J.J.Bausch, D.C.Zenger, R.Ludwig & R.D.Sisson (1993) Using
Multiple Expert Systems with Distinct Roles in a Concurrent Engineering System for Powder
Ceramic Components. Proc. AI in Engineering Conference.
K.J.Werkman (1992) Multiple Agent Cooperative Design Evaluation using Negotiation.
J.S.Gero (Ed.), Artificial Intelligence in Design ’92, KluwerAcademicPublishers.
49
In: