From: AAAI Technical Report SS-02-02. Compilation copyright © 2002, AAAI (www.aaai.org). All rights reserved. Congregating Christopher and Market Formation H. Brooks and ~Edmund H. Artificial Intelligence Laboratory University of Michigan Ann Arbor,MI 48104 {chbrooks, durfee}@umich.edu Abstract Durfee asauctions or coalitions canbe usedwithin eachcongregation to decide whichagentsshould perform particular Agents in a multiagent system arenottypically entasksor exchange goods. tirely self-sufficient; instead, theyfrequently needto Congregations area useful self-organi~.ing mechanism enlist otheragents to perform tasksforthemor to for several reasons. First,theyallowthesystemto exchange goods or services with them. This creates scale:oncean ~agenthasjoineda congregation, decia problem: howcan an agent efficiently locate other sionsas to whomit shouldinteract withcan(atleast agents to workor trade with? As the numberof agents grows, the cost of this computation can becomeproinitially) consider onlytheothermembers of itsconhibitively large. Onesolution to this is for the system gregation. As longas thecongregation sizeremains to self-organize into smaller groups of agents. In this relatively constant, thepopulation can growwithout paper, we apply the idea of congregating to a model impacting an agent’s decision process. Second, it proof an information economy.Weillustrate howparticvidesthesystemwitha typeof institutional memory. ipants in this economycan self-organize into a set of Thecongregation canbe considered as an entityunto marketssuch that agents are able to find suitable partitself, muchinthewaythatwecantalkaboutan anthill ners while retaining low computationalcosts. Weshow or a football teambothin termsof itsmembers andas howcongregatingcan help allocation problemsscale to a singleobject. Theexactpopulation of a congregation large populationsby allowingagents to interact locally. maychange as particular agents leaveor arrive, andyet the groupas a wholeis ableto maintain a long-term Introduction existence. Third, it canserveas a wayof coordinating agents’ decisions as to whichgroup(s) to join.By deIn a dynamic, openmultiagent system,agentshaveto scribing a congregation in a meaningful way,a system continually makeandre-evaluate theirdecisions as to designer canhopeto attract agents of particular types whichotheragentstheyshouldinteract with.These or interests. interactions mightincludethebuyingandsellingof Congregating canbe seenas a formof ’cooperative’ goods,negotiation overa setof tasksto perform, or learning, or group-level learning, eventhough theparcooperation to achievea jointgoal.A greatdeal ticipants are all completely self-interested. The populaof research in bothgametheoryandartificial inteltionas a wholeistrying tolearnhowto organize itself. ligence, cooperative gametheory(Mas-Colell, WhinIn mostmultiagent systems, the agentsareplayinga ston,& Green1995)andcoalition formation (Shehory general-sum game,in whichthereareparticular config& Kraus1998),auction theory(Wellrnan et al.2001), urations of congregations that will lead to higher utility and teamformation (Tambe1997)hasfocusedon how fora largefraction of theagentpopulation. Therefore, groupof agentscanmakethesesortsof decisions. Two of thesecongregations canbe viewedas an problems typically arisein thesesortsof approaches: discovery adaptive distributed search through the spaceof allposscaling to domains with large numbers of agents, and siblecongregations, in whicheachagenthasa partial dealing with the fact that the composition of the agent influence on theexploration process. population is not typically knownto any single agent. In the rest of this paper, we willdiscuss theroleof Our research addresses the problem of how a large congregations in an information economy, d’efine more population of agents can, in a self-organizing fashion, precisely what we mean by a congregation and describe separate itself into subgroups knownas congregations. it in termsof a simple modelofan information economy. Ideally, each congregation should contain a collection of We will then show experimental results that suggest agents whowould like to interact with each other. Once that congregations can improve the overall utility of these groups exist, standard allocation algorithms such the agentsin a multiagent system.We willconclude withfuturedirections forstudying theformation and Copyright (~) 2002, AmericanAssociation for Artificial Indynamics of congregations. telligence (www.aaal.org).All rights reserved. 33 Congregating in Information Economies through auctions, each of which was described using an ontology which indicated the goods being bought and sold. These auctions were instantiated via the Auction Manager Agent (AMA), which would dynamically create new auctions when it noticed unsatisfied consumer demand and delete auctions when they experienced a lack of activity. This was a centralized process which allowed for the partitioning of the market space into separate congregations, makingthe location of potential customers or goods a manageable process for the agents involved, as well as the humanusers. In our exploration of congregating and market formation, one goal is to reproduce this process without the intermediation of a centralized market maker (the AMA).Instead, we propose that this functionality can be handled by a distributed set of market makers, each of which has the goal of attracting buyers and sellers such that its market is successful (measured in terms of either number of trades or surplus generated). In essence, we would like for the formation of multiple markets, each of which is composedof buyers and sellers with complementary interests, to be an emergent phenomenon. The purpose of this work is to identify whether congregating can in fact be a viable strategy for a large group of self-interested agents. Our past research (e.g. (Brooks et al. 1999; Brooks, Durfee, & Das 2000)) has focused on how producers an information economycan locate consumer niches and efficiently learn the preferences of consumersin these niches. It is this first problem, that of locating a niche, that is most relevant to congregating. In this problem, two or more producers are each separately selecting a setof goodsto sellanda pricing mechanism forthese goods.In doingthis,theyseparate theconsumer populationintooneor moremarkets, eachof whichbuys fromoneproducer, plusthemarketconsisting of consumerswhochoosenot to buy fromany producer. Each producer’s problem isthentofinda profitable location in price/product space,subject to thedecisions of the otherproducers. In thispaper,we taketheproblem of marketformationonestepfurther, anddrawdirectly on thecongregatingmetaphor to examine theprocess of marketformationin a multiagent systemcontaining a largenumberof producers andconsumers. Elsewhere (Brooks Durfee2002),we present a general modelof congregations,focusing on theaffinity groupdomain. Thiswork showed how congregating can be difficult when the solution space is very sparse, and suggested the introducCongregations tion of labelers as a signaling mechanismthat served to In this section, we define more precisely what we mean coordinate agents’ decisions. It also showed how congregating could be used to improve average payoff when by a congregation, both generally and in terms of our agents with heterogeneous preferences played a coordiinformation economy model. A more detailed presentanation game. Our current paper applies congregating to tion of this model(with minor differences) can be found information economies. Weshow how congregating can in (Brooks& Durfee 2002). be used to stimulate producers and consumers in an inWe beginby considering a generalmodelof a mulformation economyto self-select into separate markets tiagentsystem.Let A = {al,a2,...,a,~) be a set of so that the average profit in the system is increased. agentsin thesystemat timet. (Foreaseof presenThe effectiveness of this approach depends upon the tation,we willassumediscrete time.However, nothnumberof markets available, as well as the specificity ingin ourmodelrequires it.)Theessential criterion of agent preferences. forcongregating to be an interesting problem is that eachagentneedsto interact withsomeothersubsetof In manylarge-scale economies, the problem of whoto theagents inthepopulation. Inutility-theoretic terms, buy or sell from is a significant one. Mechanismssuch each agent a has a utility function U~ : P(A) -+ as auctions which pair buyers and sellers can be used to (P(A) is the power set of all agents.) This function do this when the number of agents is relatively small, dicates the utility that agent a receives from interacting but if the population is large or the auction is combiwith a particular set of other agents. natorial, both the allocation of goods and the selection of bids becomes a difficult computational problem. A It is assumedthat an agent cannot interact with every other agent, due to computational and communication solution to this is to construct several smaller auctions, each of which contains buyers and sellers with ’similar’ constraints. (Recall that n is large.) Therefore, at every interests. time t an agent will need to choose a subset of agents to interact with. One inspiration for this approach was the UniverCongregations are a way of simplifying this decision sity of Michigan Digital Library (UMDL) (Duffee al. problem; rather than choosing agents from the entire 1998), which was a market in which automated agents population of size n, an agent can join a congregation would buy and sell information goods. The UMDL contained consumers with manydifferent interests, and of size c <~ <: n. This allows the agent to reduce the size of its search problem by considering the c agents in its providers offering a wide set of different types of articles. Each consumer had a User Interface Agent (UIA) congregation to interact with. Additionally, in domains in which agents have symmetric interests (meaning that that located sellers of relevant information goods. The sellers were also represented by agents, knownas Col. if agent a wants to interact with agent b, then b wants to interact with a), forming congregations will improve an lection Interface Agents (CIA). These agents interacted 34 agent’s chances of a ’successful’ interaction byallowing agents to grouptogether withotherlike-minded types. Mostgenerally, a congregation c is simplya tuple < l,Ac,t >,where Acisa setofagents {ai,...,aj}who haveallcollocated at a ’location’ l at timet. (This neednotbea physical location; it couldbea particular multicast address, radiofrequency, or mailing list.The pointis thattheycommunicate onlywithotheragents in thecongregation.) We assumeforsimplicity thatan agentis onlya memberof onecongregation at a time; futureworkwillrelaxthisassumption. A congregation is defined intrinsically by itsmembership. It maybe usefulforthemembers of a congregation or forother agents, suchas market makers, to tryto describe a congregation e~-~rinsically by characterizing somequality shared byitsmembers, butthisisjusta label. Letus makethisallmoreconcrete by placing it in termsof an information economy. Our setA of agents consists of twodisjoint subsets: a setB ofbuyers anda I Ourcommodity setS of sellers. is information goods: an information goodis an article of a particular categoryc E {ci,...,ck}. Examples of categories include Sports, Arts,International News,andso on.At each timet, a seller is ableto select a typeofgoodtooffer.Eachbuye has a category c* of goodthatit most r prefers; articles inthiscategory havea reservation value r. A goodfromanother category c’ is valuedat: other, or a morecentralized mechanism suchas an auction,wherebidsaresubmitted to a central authority whichcomputes allocations. In eithercase,thereis somecomputation thatmustbe performed in orderto determine an allocation. In thecaseof Contract Net, thiscomputation is the exchange of messages, and in thecaseof an auction, it is thewinnerdetermination algorithm. In eithercase,thiscomputation doesnot comeforfree;we assumethatthe agentsin the systemmustpayforit.By separating intoseparate congregations, theycaneasetheircomputational burden, hopefully without a largedecrease in theefficiency of theallocation of goods. In fact,thegoalof theexperimentsin thispaperis to determine how muchgross efficiency is lost,howmuchcomputation is saved,and theresulting netcostsas thenumberof congregations isvaried. In orderto capture thisnotionof agentsseparating intosubgroups, ourmodelalsoconsists of a set M = (m1,...,mq}of markets.Theseare the locationsdescribed above;buyersandsellers willcongregatehere.Eachbuyersandsellerwillchooseexactly one marketfrom M to join.At a time t, the marketwillcloseandanallocation willbe computed forall agents inthemarket. Thecostofthisallocation willbe sharedequallyamongstallagentsin the market.The mechanism we useis a standard Nth-price Vickreyauction. This was chosen for simplicity; since all goodsare -dl privately valuedandthereis no resale, thedominant V(c’)= r(l [c* k- ) (I) strategy is foragents to bidtheiractual valuations for Notethatthisassumes thatcategories arearranged acarticles (MacKie-Mason & Varian1994).However, any cording to a ’similarity’ metricon thek axis.While allocation mechanism canbe usedin thismodel. If costs thisisa simplification, itissufficient forourneeds. It aresuperlinear in thenumber of agents, thesamequalalsoprovides ourmodelwithsomesimilarities to the itative effects willbe seen.EachbuyerbidsitsreserHotelling model,whichis a commoneconomicframevationvalueandthepricesandallocation of goodsis workforproduct differentiation. Hotelling models typcomputed. Ea~ agentcan thendecidewhetherto reicallycontainone or moreproductdimensions along mainin thiscongregation or leaveforanother; thisdewhichproducers candifferentiate themselves, andconcision willbebasedon whether thevaluereceived (less sumerswithheterogeneous preferences alongthesedicomputation cost)exceeds theagent’s threshold. mensions. Eachproducer mustdetermine whereto loClearly, ina single market, theallocation whichmaxcate,giventheexpected locations of theotherproducimizese~iciency canbe computed. Thiswillbe usedas ers.Anderson, et al. (Anderson, de PMma,& Thisse a benchmarkin our experiments. However,comput1992)providea thorough overview of Hotelling-style ingtheoptimal allocation in a combinatorial auction is models. NP-complete (Sandholm 2001),whichmeansthatthe In ourmodel,eachagenta hasa threshold ~-~which computational costswillbe prohibitively largeas the indicates thefraction of r it wantsto receive froma numberof agentsin a marketbecomeslarge.Evenin transaction. Thisthreshold canbeinterpreted either as simpler allocation problems, suchas coalition formaa production_or consumption cost,or as a formof saristion,the numberof messages exchanged is polynomial ricing, whereby alltransactions yielding utility greater in thenumberof agentsin thesystem, whichleadsto than~’ax r areconsidered satisfactory by agenta. scalability problems as thenumberof agentsbecomes In a simpleworld,thiswouldbe allthatis needed large. fora model;buyersandsellers couldbe pairedup usElsewhere (Brooks& Durfee2002),we haveanalyzed ingsomesortof predetermined mechanism, eithera detheconvergence properties for a simplecongregating centralized mechanism suchas Contract Net (Davis model, focusing on thelength of timeneeded fortheopSmith1983),in whichagentsbroadcast offersto each timalsetof congregations to form.Thatworkshowed thatcongregating canleadto highernetutility in doIWeassume’that there isnoresale; buyers consume their mainswhereagentstryto locateotheragentsof the goodsimmediately. Economies containing supply chains are same"type".One observation fromthatworkis that alsoa promising areaofapplication forcongregations, but congregating isuseful because ithelpsto builda "critarenottreated here. 35 icalmass;" a groupof agents willforma congregation Average pmllti:er narstienis number of nvukem immNmn thatis mutually beneficial, andsothecongregation will persist overtime.Thisprovides a focalpoint fortheat300 traction of otheragents thatmaywantto be a partof thiscongregation. Essentially, thecongregation allows 250 a subsetof agents to "holdstill"andallowothersto findthem. Inthispaper, weareexplicitly interested inthelearn|,. ingdynamics of marketformation: whatdo thetransitional congregations looklike,howareagents’ aggre10o gateprofits affected by thecongregating process, and whether congregating canimprove overall profit by pro5O , ,, , ,In" viding thissortof critical mass.Dueto thecomplexity of the problem andour desireto examinethetransi[ , , .. -, 0 , , .... 0 ’ 100 1(]oo tionalbehavior of thesystem, we focuson experimenNurn~lra( .qm~mw talmethods to determine whencongregating is a viable strategy, andhowoverall profits (asa measure ofsysFigure 1: Cumulativeprofit per iteration over all sellers, temperformance) changeas the numberof marketsis averaged across 10 runs of 1000 iterations each (x axis altered. is log scale) P.~mputJ~nJ ~ m numb~ el ~ ~- Inemam~ Using Congregationsto Improve Net Profit 4O0 In thissection, we present experiments demonstrating theusefulness of congregating whenagentsmustincura computation costforcomputing an allocation of goods. We showhow congregating is a usefulstrategywhen agentsmustpay forcomputation, andstudytheefliciencyof theallocation (bothgross,pre-computation costandnet,post-computation) as thenumberof marketsisvaried. In thefirstexperiment, we consider an information economycomposedof 50 producers and 50 consumers. Thereare 10 categories of information goods;at the ’ 10 1~ beginning of the experiment, eachproducer randomly chooses a category of goodto sell.Eachconsumer also Figure 2: Cost of determining the optimal allocation hasa favorite category (drawnfroma uniform distribution) andpreferences overnon-favorite goodsindiof goods as the number of markets increases (avercatedby equation1. Reservation valuesfor a conaged over 50 instantiations with 50 producers and donsumer’smost-preferred goodaredrawnfromU[5,10]. sumers). Thenumberof markets is fixedat thebeginning of the experiment; it is assumed thatallagentsknowm, the numberof markets, andallagentshaveagreedto share of congregations is varied whenagents exhibit somesort equally in thecostsof computation in determining an of adaptivebehavior.Randommovementis meantto allocation. We assumethatcomputation hasa costof provide an approximation of "typical" agentbehavior 0.1 foreachmessagesentor comparison made.Since whenthepopulation is large.) we areusinga generalized Vickrey auction to compute allocations, it isstraightforward forthemarket to hold FigureI showstheaverage cumulative profit achieved an auction, determine thecostof thiscomputation, and by allsellers (averaged across 1000iterations) asthe charge eachagentaccordingly. numberof markets is increased. As thenumberof marAftereachmarketclosing, everyagentis ableto ketsisincreased, initial profits aremuchlower, as it becomesmoredimcultforeachagentto locatea conchange markets. If theproducer receives lessthan~"x r gregation in whichit canfinda suitable match.Also, foritsgood,,orif a consumer receives a goodthatit values at lessthanT X r, itwillmovetoa newcongrethe systemas a wholereachesa suboptimal configugation chosen at random. In ourinitial experiments, ~ration. However, thisfiguredoesnottakeintoaccount wassetto 0.5forallagents. Thisisvaried in section. thecostsof computation. Figure2 showshowcomputational costsfalloffquickly whentheeconomy contains (Notethatwe are not advocating randommovement as an optimal agent strategy. Instead, we are interested in morethanone market.Thesetwographsare combined characterizing the behavior of the system as the number infigure 3.Inthisfigure, wecanseethat, eventhough a 36 300 . .o , Average NetPmnl perItemllon , , .... ~ ...... ~.-.. I" ¯ "" ........... , " , .............. 25O ~....-r..~,,~.~..rrr:..’:’:’.’. ....................... 2OO 2OO ! Z/ 5O 0 .5o 0 0 i 100 i 200 i 300 i 400 i SO0 i (~0 l ?00 i 800 gO0 1000 , q~ , , ..... i tO , NumbMof Figure 3: Net cumulative profit per iteration over all sellers, averaged across 10 runs of 1000iterations each. single market is able to computethe optimal allocation, the computational costs are high enough that producers end up with a negative net profit. As the number of markets is increased, the computational cost falls off muchmore quickly than profit, until at 25 markets the average cumulative net profit is maximized. As more markets are added, net profits begin to decline, as the computational savings is outweighed by the difficulty of finding a market cont~.in!ug a suitable matchup. In figure 3, we can see the dynamics of market formation as the number of markets is increased. With a very small number of markets (3 or 5) a stable configura-tion is reached almost immediately. As we increase the number of markets to 10 or 25, convergence is slightly slower, but the configuration of congregations that is found yields higher average profit. As the number of markets is further increased, the quality of the configuration falls off slightly, and convergence time becomes muchlonger. It is difficult to quantify this comparison directly from figure 3, due to the number of lines and the fact that total performance is actually the area under each curve. Figure 4 makes this tradeoff more explicit; it showsthe average net profit (over 1000 iterations) as the number of markets is increased. Here we can clearly see that net profit increases steeply as the numberof markets is increased, and then falls off after m--25. Wecan also see that introducing too many markets, as in the case of m = 250 or m = 500, introduces too muchinefficiency without a significant extra reduction in computational cost. Buyers and sellers spend a great deal of the initial iterations in a marketwith no suitable agents. In our next experiment, we examine whether the advantages of congregating scale as the number of agents jn the system is increased. Wevary the number of producers and consumers from 10 to 500 of each, in each case setting the number of markets to half the number 37 , ¯ 100 , - . , , ~ Figure 4: Average net profit (over 10 runs of 1000 iterations each) with 50 producers and 50 consumers as the numberof markets is varied. (x axis is .log scale) 1o o l I I00 I 180 l ~00 I ~mo I ~ I ~ I 400 l 450 ~0 Figure 5: Net profit per iteration per producer as the number of producers (and consumers) is varied between 10 and 500. of producers, and compare the net (post-computation costs) per producer. The average profit per producer over the course of the run is presented in figure 5. As we can see, congregating is in fact a method that is able to scale to large numbersof agents. Profits are invariant as the number of agents in the system is increased. (In fact, net profits for small numbersof agents are slightly lower; this is an artifact of the wayin which agents are randomly generated. For low numbers of agents, it is less likely, for any given producer, that there is a consumer whose most desired good corresponds to that produced.) As long as the system designer has rough idea of how many other agents are in the economy and is able to construct the appropriate number of markets (half the numberof producers in this case), market size remains constant, the economy can scale, 30o m 28o 1I 180 leo Figure 7: Net profit per iteration as the threshold is varied between 0.1 and 1, averaged over 10 experiments. r is the agent’s reservation value. X-axis is log scale. Figure 6: An illustration of the changes being performed on a simplified landscape. The top figure shows a simple two-peaked profit landscape. As agents’ r is increased, the "sea level" increases, lading to the second figure, the original landscape is indicated by a dotted line. Whenconsumers reduce the number of categories they value, the result is the attenuated landscape shown ¯ at the bottom. The heights of the peaks remain constant, but the area of the plateaus increase. and producers and consumersare still able to find each other and make successful transactions. Of course, these results are dependent on the particular costs and reservation values chosen. The higher the ratio of reservation values to computationcosts, the more utility will be placed on an optimal solution. On the other hand, as computation costs become a significant portion of net profits, agents will prefer cheaper solutions and congregating will gain appeal. Varying Consumer Preferences A largepartof theeffectiveness of congregating comes fromitsability to serveas a coordination mechanism foragents’ decisions. A subset ofagents willmoveinto a particular market and,if theyarehappy,staythere. Thisreduces thecomplexity of theproblem foragents whohaveyetto finda suitable market,sincefewer agents aremoving. If we visualize thecongregating processas a search overa landscape, congregating allows thesearchprocess to movealonga gradient moreeffectively. Of course, thisassumes thatthelandscape has a gradient. In thissection, we alterthegradient ofthe landscape andexamine theeffectiveness of congregating. 38 In our first experiment, we alter agents’ threshold 1-. Recall that r is the fraction of reservation value an agent must receive to be satisfied with a transaction. As ~- increases, an agent is happier in fewer congregations. In terms of the search landscape, this lowers all the peaks equally, as if there was a flood. Fewer points on the landscape have positive profit, but the topology is not deformed. Figure 6 illustrates this. Experiments were again conducted with 50 consumers and 50 producers. All experiments were performed with 25 markets, seen to be the optimal number in our previous experiments, r was varied between 0.1 and 1.0. Results are shownin figure ? As we see from figure 7, when7" = 1, the process takes longer to converge to a solution, but the quality of the final solution is improved. Whenr is small, the process converges more quickly, but to a local optimum. The encouraging result here is that even when the threshold is high, meaning that agents have a high cost of producing or consumingarticles, congregating still quickly finds a desirable solution. In our second experiment regarding consumer preferences, we alter the number of categories each consumer values positively. Wedo this by giving each consumer a value n which indicates the number of categories they value positively. If an article offered by a producer is of a category less than ~ away from the consumer’s preferred category, it is valued according to 1. Otherwise, it has value 0. This has the effect of attenuating the congregating search space. The gradient around an optimum is increased, producing steeper peaks and larger flat areas. While the profits attained by the system at the optima do-not change, the profits for nonoptimal configurations of congregations decrease, since consumersvalue fewer categories. In essence, the search problem becomes more di~cult. This is also illustrated in figure 6. . - - i ~#~f~2 ~ .~ ~-~. ~.~ otheragents inthesystem; it cansimply concern itself withtheotheragents in itscongregation. "" .’~’ ~ We havealsoshownthatcongregating is a technique whichallowsrnultiagent systems to scaleup to large numbersof agents.When agentsmust considerthe costsof computation in determining who to purchase fromor howto allocate goods,congregating allowsthe system toscalesothatnetprofit remains relatively constantas thesystemgrows. / / / I00 .~÷,@. / ..,o .... n-5........ ~- /.S ........ ,o ’ ’ ...... i~o ..... 1000 tlm’mlon Figure8: Netprofitperiteration perproducer as the numberof categories (outof a possible 20)valfied eachconsumer is varied. Results areaveraged over10 experiments. X-axis is logscale. Additionally, we haveshownthatcongregating isrelatively robustto theshapeof agentpreferences. Even whenconsumers havea veryhighprocessing cost,or verynarrow preferences, congregating stillallows producersandconsumers to makesatisfactory exchanges, although thetimeneededto discover theoptimal con° figuration ofcongregations increases. One question that must be asked is the applicability of these results to other domains. Our experiments rely on the assumptions that agents have utility functions which they are able to evaluate exactly, that agents knowtheir preferences, and that agents have someprefOnceagain,experiments wereperformed with50 conerence over who they choose to associate with. Many sumers, 50 producers, 25 markets and20 potential cateconomic domains meet these assumptions, but there egories, r wasfixedat0.5.Results forthisexperiment are other domains in which agents cannot easily exareshownin figure 8. press their utility functions. In addition, we have not’ As we canseefromfigure 8, as consumer preferences touched upon the problem of market formation and adbecomemorespecific, thesearchprocess becomes more vertising, apart from declaring that markets exist and diflicult; manyconfigurations of congregations do not that all agents know about them. A crucial problem contain a critical massof agentsreceiving a positive in the UMDL was the description and dissemination of profit. In termsof thesearchlandscape, thesizeof information about these markets , so as to encourage theplateaus increases. Themoreinteresting resultis the congregating of the "right" agents. This remains a that,eventhoughthesearch process becomes morediftopic for future research. ficult, through theuseof congregations, theagentsin There are many other potential directions for this thesystem arestillableto self-organize intocongregationswhichyielda largefraction of theoptimal profit. research. As noted previously, goods in this economy are immediately consumed. If goods are instead transThistellsus thatcongregating is a method whichis roformed and resold, then the need for well-constructed bustto theparticular shapeof agentpreferences; even markets becomes even more important. Additionally, whenconsumers haveveryspecific preferences as to the this work does not explicitly treat nonstationarity in goodstheywant,congregating allowsproducers and the consumer population. An important question to ¯ consumers to locate eachother. ask would be whether the existence of congregations makes consumer and producer entry more di~cult. Discussion In thispaper,we haveprovided an example of howcongregating canbe applied to information economies and usedto modeltheprocessof marketformation. As we haveseen,if thep~ticipants in a marketmustpayfor theircomputation, theyareoftenhappier to realize an allocation whichyields slightly lessutility, butcanbe foundat a muchlowercomputational cost.Themechanismisessentially a self-organizing one;initially, agents whicharehappyin a market willstaythere, providing a fixedlocation thatother’like-minded’ agents canfind. The onlycommonknowledge thatis requiredis that theagents allagreeontheexistence ofa setofmarkets at thebegivni-g of theeconomy’s lifetime. Oncethis initial commitment is made,an agentneednotconsider theglobal stateor worryabouttheidentity of allthe 39 Finally, weearlier stressed thenotion ofa congregationasan entity untoitself. Thisleadsoneto consider notions of groupselection, whereby it becomes rational foragents within a congregation tobehave altruistically, soas to increase thefitness of therestof thecongregationandimprovetheirchanceswhencompeting with agentsoutside of thecongregation. In environments in whichagentstypically interact withina congregation, butoccasionally encounter agentsfromthepopulation atlarge, thiscanbea veryeffective strategy (see(Wilson1980)forecological examples). Thiswouldalsolead us to examine therelationship between individual and group-level learning: as individual agentsdevelop more sophisticated strategies, howdoesthecomposition of thecongregation change? Acknowledgments The authorswouldliketo thankJeffMacKie-Mason andBobGazzale fortheircontributions tothisresearch. Thisworkwassupported in partby NSF grantsIIS9872057 and IIS-0112669. References Anderson, S. P.; de Palma,A.; and Thisse,J.-F. 1992. Discrete Choice Theory of Product Differentiation. Cambridge, Massachusetts: MIT Press. Brooks, C. H., and Durfee, E. H. 2002. Congregation formation in multiagent systems. Journal of Autonomous Agents and Multiagent Systems. To appear. Brooks, C. H.; Fay, S.; Das, R.; MacKie-Mason,J. K.; Kephart, J. O.; and Durfee, E.H. 1999. Automated strategy searches in an electronic goods market: Learning and complex price schedules. In Proceedings of ACMEC-g9, 31-40. Brooks, C. H.; Durfee, E. H.; and Das, R. 2000. Price wars and niche discovery in an information economy. In Proceedings of ACMConference on Electronic Commerce (EC-O0), 95-106. Davis, R., and Smith, R. G. 1983. Negotiation as a metaphor for distributed problem solving. Artificial Intelligence 20:63-109. Durfee, E. H.; MuUen,T.; Park, S.; Weinstein, P.; and Vidal, J. 1998. The dynamics of the UMDLservice market society. In Klusch, M., and Weifi, G., eds., Cooperative Agents II, Lecture Notes in Artificial Intelligence. Springer. 55-78. MacKie-Mason, J. K., and Varian, H. 1994. Generalized Vickrey auctions. Technical report, University of Michigan. available at: http://wwwpersonal.umich.edu/jmm/papers/gva3.pdf. Mas-Colell, A.; Whinston, M. D.; and Green, J. R. 1995. Microeconomic Theory. NewYork: Oxford University Press. Sandholm, T. 2001. Algorithm for optimal winner determination in combinatorial auctions. Artificial Intelligence, to appear. Shehory, O., and Kraus, S. 1998. Methods for task allocation via agent coalition formation. Artificial Intelligence 101:165-200. Tambe, M. 1997. Toward flexible teamwork. Journal of Artificial Intelligence Research 7:83-124. Wellman, M. P.; Walsh, W. E.; Wurman, P. R.; and MacKie-Mason, J. K. 2001. Auction protocols for decentralized scheduling. Games and Economic Behavior 35(1/2):271-303. Wilson, D. S. 1980. Natural Selection of Populations and Communities. Menlo Park, California: Benjamin/Cummings. 4O