From: AAAI Technical Report SS-01-03. Compilation copyright © 2001, AAAI (www.aaai.org). All rights reserved. Toward Flexible Trading Agents Peter R. Wurman North Carolina State University Suite 110, Venture1; Centennial Campus Raleigh, NC27695-7535 USA wurman@csc.ncsu.edu Abstract Themajorityof studies of agent decisionmakingin environments with multiple auctionsassumehomogeneity in the rules of the auctions. However, online marketsare much morefragmented, and productsare often for sale in a variety of venueswithdifferent rules. Wepresenta conceptual approachto designinge xible trading agents that is broad enoughto encompass both multipletypes of auctionsand a variety of user preferences.Thegeneralizationsdescribed lead to natural representationsof the agent’sdecisiontask as either a Markov decisionprocessesor an extensiveform game,dependinguponthe formof the marketmodel. 1 Introduction Electronic auctions have rapidly becomeone of the most vibrant and widespread applications in e-commerce.Each day, consumerslist millions of items in electronic auction sites suchas ebay, and it is dif cult to nd an industrythat is not launching a vertical B2B"exchange".This profusion of auctions poses newchallengesfor the participants in the marketplace. Currently, if one does a search for a common product, such as a PalmPilot, one will nd there are hundreds of active listings hostedon a dozendifferent Websites each with potentially different auction "types". In addition, each auction has a unique closing time, and can have other attributes whichaffect the buyersinterest level, suchas text describing the quality of the product or a measureof the seller’s reputation. For a variety of reasons, the actual Internet marketplace has evolvedinto a fragmentedmarketin whichrelated products are sold in various auction formatswith subtle but very importantdifferences in rules. Seeminglyminordifferences can induce radically different bidding behavior on the part of the bidders. Take, for example,the different strategies induced by the ascendingauctions that end after a period of inactivity in whichno newbids are received (eg. Amazon.com),and those that end at a xed time (eg. ebay). latter set of rules inducesparticipants to "snipe",that is, to hold off bidding until the nal moments, and then try to submit a highbid just before the auctioncloses. Thisstrategy is ineffective in auctions withinactivity-basedclosing rules. Copyright©2001, American Associationfor Arti cial Intelligence(www.aaai.org). All rights reserved. 134 Nolicense: PDFproducedby PStill (e) F. Siegert - http://www.this.net/~frank/pstill.html Makingbidding decisions in fragmentedmarketplaces is complex;to our knowledge,people handle this problemby simplyrestricting their attention to a small subsetof the auctions of potential interest. Clearly this can lead to suboptimal outcomesfor both the buyers and the sellers. Software agents havethe potential to greatly enhancean individual’s ability to effectively participate in these environments.We call such software programstrading agents. In somesituations, such as consumerpurchasing on the Internet, the conditionsthat a real agent will face will vary from episode to episode. Evenwhenmarket conditions are more consistent, as might be found in a B2Bexchange, we need to be able to quickly develop and deploy agents in each specialized environment.Adoptionof agent technologies will be limited if an agent’s bidding strategy must be hand-coded for each marketplace, and then recoded when the rules of the market change. Thus, we argue that trading agents mustbe designed to be exible, that is, capable of perceiving both the context and the dynamicstate of the marketplacein order to makerational decisions. Consider the decision problem facing an agent whenit desires to purchasetwounits of a particular object for sale. The agent searches the web and nds the following three 1auctionslisting the desiredobject. ¯ AuctionA lists a single unit. The auction rules are those of a typical ascending auction with axed closing time, designatedta. ¯ AuctionB lists a single unit. The auction is also an ascendingauction, howeverit will close after time tB when ~Btime passes without any newbids. ¯ AuctionC lists ve units. It uses a type of multi-unit auction in whicheach bidder pays the price of the lowest accepted bid. Indivisible bids are accepted. This auction closes after time tc when6c time passes without any new bids. The agent could satisfy its goals by purchasingtwo items in auction C, one in C and one in A or B, or one in each A and B. The bidding strategy that accomplishesthe selected goal is dependentuponthe expectednal prices in each auction, the rules of eachauction, and the relative closing times IAll threeof the auctionrules describedare widelyusedonthe Intemet. of each auction. Moreover,this decision is not static--the agent’s expectationsof the nal prices dependsuponthe current price quotesand the actions of the other participants. The research to date, both in economicsand computerscience, viewsthe multi-unit or multi-object allocation problems from one of two perspectives. Onebodyof work takes the perspective of the designer of the mechanismwhose goal is to select a set of rules that maximizessomeobjective. The approachesstudied include both auctions for sets of identical objects (Demange,Gale, & Sotomayor 1986; Gale 1990; Rodr’guez et al. 1997) and combinatorial auctions (Lehmann, O’Callaghan, & Shoham1999; Parkes 1999; Wellmanet al. to appear; Wurman&Wellman2000)Thesecondperspective is that of the participant. This bodyof workfocuses on decision makingfor scenarios in whichthe auctions are homogeneous (Boutilier, Goldszmidt, & Sabata 1999b; Hu &Wellman1998; Park 1999; Rust, Miller, &Palmer1993). Weare interested in decision makingfor heterogeneousmarketplaces. This paper presents a conceptual approachtowardbuilding exible trading agents. In Section 2, we describe the architectural elements of a exible trading agent and suggest someuseful formalisms. Section 3 describes someof the issues related to gametheoretic and decision theoretic strategy selection. 2 Components of a Trading Agent Weuse the term marketplaceto refer to the combinationof the user’s agent (called the principal agent), the auctions, and the other bidders in the system. Similarly, the decision makingpolicy ofa exible trading agent can be viewedas a function with three inputs: 1. A modelof the user’s preferences. 2. Alist of auctionsthat are relevant to the task and a method of determiningthe particular rules of each auction. 3. Amodelof the other bidders in the auctions. Fromthese inputs, the agent mustderive a bidding strategy that whenexecuted (ideally) maximizesthe user’s expected payoff. TheStrategy GenerationEngine(SGE)is responsiblefor derivinga biddingstrategy. Figure1 illustrates howthese componentst together. The following discussion identi es someissues associated with each of the four components. 2.1 User Preferences Eliciting the user’s economicpreferencesin a mannerthat is not overly burdensometo the user and yet acquires enough information to be of value is a challenging problem. For example, the agent should not ask the user howmuchshe values every make/modelof computermonitor that is currently being sold online. Instead, wewouldlike the agent to ask a fewquestionsfromwhichit can infer a suf ciently completepreference structure. Althoughpreference elicitation is an interesting area of research, for the purposesof this paper we assumethat we havethe users’ preferencesin the formof a utility function. Weadmita variety of preference structures, including both 135 Nolicense: PDFproducedby PStill (c) F. Siegen- http://www.this.nel/~frank/pstill.html U set Pref~renGes A uctbn Rules M azket Model st~at~y G en~t:bn Eng~he ] ~ s~gy Figure1: Anarchitecture for trading agents. substitutable and complementary preferences, and both independentprivate values and af liated common value problems. 2.2 Auction Rules In previous work (Wurman,Wellman,& Walshto appear), we have described a parametrization of auction rules that encompassesthe majority of common auction formats, including those used by most Intemet auction sites. Moreover, the parametrizedrules are instantiated in an online auction server called the Michigan Intemet AuctionBot (Wurman, Wellman,& Walsh 1998). The parameters are organized according to whether they governthe admittanceof bids into the system,the revelation of informationby the auction, or the policies and timing of eventsthat clear the auction. For example,rules associated with bid admittancedeterminewhois allowed to bid, the format of bids, and whether and in what mannerbids must "improve".Rulesthat control clearing specify the eventsthat trigger a clear (e.g. at axed time or after no bids are received for z minutes), and the policies used to computethe prices, quantities, and trading partners involvedin the resulting exchanges. The rules are, to a large extent, orthogonal, enabling a combinatorial numberof different auction types. Not only can the vast majority of auctions available on the Internet can be de ned using the parametrizedrules, a large number of newand potentially useful auctions can be identi ed. The AuctionBotserver also includes an Agent ProgrammingInterface (API)whichpermits software agents to interact with the server. Throughthe APIthe agent can request the speci c rules of each auction. Currently, the AuctionBot returns a text string specifying the auction rules. Work is currently underwayto convert the API to an XML based interaction. 2.3 Market Models Howmuchand what type of knowledgeof the other participants in the marketdoes an agent need to perform effectively? This is one of the fundamentalresearch questionsin the design of trading agents. Amongthe choices of market modelsare: I. A modelof the evolutionof prices that ignores the behavior of individual non-principalparticipants. This essentially treats the external marketas a single entity wecall a representative agent, with (potentially uncertain) reactions to the agent’sbids. A modelof the individual participants in the auctions. In . general there are a variety of issues that comeup when modelingother participants. ¯ Are the other agents’ preferences known? . Are the other agents’ strategies known? ¯ Whatinformation do the other agents have about our agent, and howis that informationused? The model used by Boutilier, Goldszmidt, and Sabata (1999b) falls into the rst category, while the modelexplored by Hu and Wellman(1998) explores multiagent learning in the secondcontext. Theseexampleshelp illustrate the spectrumof potential models.At one end are gross level modelsin whichthe behaviorof the other agents is encapsulatedin a single representative agent. At the other end of the spectrum we have gametheoretic models which assumeperfectly rational, self-interested agents with common knowledge of each other’s utility functionsand capabilities. Theapplicability of the potential representations is also affected by features of the particular market.For instance, consider a marketplacewith several sequential auctions for individual, substitutable objects and axed numberof bidders. As the marketplaceprogresses and someof the objects are allocated, the numberof agents competingdecreases. Moreimportantly, the past behaviorof the remainingagents, and the fact that they haven’t yet purchasedan object, provides information to the principal agent regarding the remainingagent’s preferences. A purely gross level representation maynot properly capture this valuableinformation. However, in practice it is quite rare to haveas muchinformationas the gametheoretic modelassumes.Eventually, we wouldlike our market modelsto be based on the types of real data that is accessibleto the agent. Manysites, such as ebay, provide identifying informationthat can be used to track the behaviorof individual participants. Weare writing softwareto collect this type of informationfroma small set of auctionsin a particular productarea andwill use it to explore the automatedconstruction of marketmodels. 2.4 Strategy Generation Engine (SGE) Theprimaryfocus of this project is to build a strategy generation engine. At an abstract level, we canseparate strategy generation from strategy implementation,that is, we assume we can generate a completepolicy a priori that species the agent’s best action for every state it mayencounter.Froma practical point of view, it is often not desirable to precompute a policy for all possible states, and therefore somereplanningis desirable. Thetask of the strategy generation engine is to compute a strategy that maximizesthe agent’s utility. The methodology that is employedin this task dependsuponthe internal 136 No license: PDFproduced by PStill (c) F. Siegert - http://www.this.net/~frank/pstill.html A gent A uctbn com pu ~ A u ctSon corn putts new sl~l~ new stmt~ responds Figure 2: Thecycle of actions. modelused to represent the inputs to the SGE.For example, if individual agents are modeled,then a gametheoretic representation maybe appropriate. Onthe other hand, if the marketis modeledas a stochastic state machine, then a MarkovDecision Process maybe a useful frameworkfor makingdecisions. In this section, we describe the SGEat an abstract level in the framework presentedby Boutilier, Dean, and Hanks(Boutilier, Dean, &Hanks1999). Weassumea discrete time environmentin whichthe agent modelsthe worldas the cycle of actions depictedin Figure2. The agent evaluates the marketplaceand selects an action (i.e. bids). Asa result of this action, the auctions compute new states. The other agents then respond to the newinformation, whichin turn causes the auctions to recompute states. Eachiteration of the cycle is called a stage, denoted t. Formally, the environmentin whicha trading agent operates is a collection of auctions, {cq,..., am}, each with its ownset of rules. The principal agent is denotedu, and the marketis inhabited by other agents, D= {dl, .. ¯, dn}. Thesubset of Dthat participates in auction i is denotedDi. If Di f’l Dj = { } for all i # ] then we can safely assume that the agent’s actions in i does not affect the behaviorof the agents in j. However,we expect that there will be many situations in whichthe participation is not disjoint among the auctionsof interest. Each auction has a state, ~rg, which is de ned by the sequence of messages(bids) that the agents have sent the auction. The marketstate is the set of auction states, S = {crt1 ... cry’}. The rules of each auction greatly in uence the state variables that mustbe stored to represent the auction. In addition, the rules de ne the agent’s permissible actions froma given state. Let fi (~r~) be the function that returns the set of actions, includingthe null action, that the principal agent cantake in auction i at timet. Typically, but not always, the permissible actions depend on only the most recent bid fromeach agent in each auction, which enables us to makeuse of the Markovassumption. However,in somesituations (eg. with activity rules) f can dependuponthe entire history of messages.In such cases, the state maybe able to be augmentedwith summaryinformationwhich captures the current restrictions imposedby the activity rules. Theset of joint actions that the agent can take at a given instant is the combinationof actions in individual auctions. That is, At = ®fi (~r~). Therules also de ne the auction’s state transition for a given action, a E A. Wedenote the state transition function for the auction i by #i. Thus, a successor state, ~r’ = #i(a, (r), is reachedwhenaction a is taken. In mostresearch on iterative auctions, this function is deterministic-if the agent bids morethan the current highest bidder, it will be the new high bidder. However, in some recent iterative combinatorial auctions (DeMartiniet al. 1998; Wurman&Wellman2000) the agent cannot always predict fromthe informationit is giventhe state transition that will result from an action. This occurs becausethe agent cannot directly observethe importantaspects of the auction’s state (i.e. all of the current bids). Instead, it can observeonlythe price informationrevealedby the auction. Whenthe state is only partially observable,the agent maynot be able to tell with certainty whichstate it is in, greatly complicatingthe decision problem. Uncertainstate transitions can also be usedto modelirregularities in systemperformance. For instance, it is a fact of Internet life that occasionallymessagesare lost in the networkand servers crash. Also, the amountof time that it takes for a bid to reach the auction server and be admitted varies with networkcongestion and server load. Thus, all bids, but especially last minutebids, incur someprobability of failing to reach the server in time to be counted. These randomevents can be accountedfor in the state transition model. Typically, the other participants will respond to the agent’s action with actions of their own.It is at this point that we start to see the interplay betweenthe marketmodel and the decision representation. If the agent is planningon using gametheoretic reasoning, it needs a methodto compute the other agents’ choicesof actions froma given state. Typically, the other agents have the samebasic choices as the principal agent, and thus the logic used in the function f can be reused. However,wheneach of the other agents is acting independently,the "marketresponds"step in Figure 2 wouldhave to be expanded. Onthe other hand, the agent maychoose to use an aggregate modelof the other participants that treats themas a single representative agent whichtakes a single (joint) action 7-. Todeterminethe representative agent’s choice(s) actions, weuse the responsefunction,/~. Whenthe response function is deterministic, our agent’s next decision point is in state o’" = gi(r, gi(a, o’)) wherer It (gi(a, o’)). Wh the functionis not deterministic,it returns a set of possible states with associated probabilities. Notethat the representative agent’s responsefunction is also constrained by the rules of the auctions. It is natural in this modelto distinguish betweenthe effects of the principal’s actions and the effects of the other agents in the system.Thus, wesuggest that a exible trading agent be basedon an explicit-event model(Boutilier, Dean, & Hanks1999). Doingso allows us to explore the approach of treating marketmodelsas moduleswith well de ned programminginterfaces. Whennecessary, we can construct an implicit modelfromthe explicit models. A bidding policy, lr is a mappingbetweenstates and actions, ~r : S --+ A. Whenthe states are only partially ob- 137 Nolicense: PDFproducedby PStill (c) F. Siegert - http://www.this.nel/~frank/pstill.html servable, the policy is a mappingbetweenobservations and actions. The purposeof an auction is to generate an allocation of the resources as a function of the messagesreceived. The allocations that an agent receivesdeterminesits nal utility. Wecall the auction’s policy for determiningallocations the allocation function, h. Essentially, the allocation function runs the samecode as the auction server, althoughit can do so withoutthe real-time constraints of the actual servers and can take advantageof certain ef ciencies in the computation. 2.5 Generating Admissible Actions The parametrization of the auction design space provides a convenient way in which to de ne and communicatethe rules of an auction. However,the decompositionof the auction rules does not correspondto a useful parametrization of the space of bidding strategies. To illustrate the point, considerthe strategies of an agent biddingin ebay’s version of the English auction--which closes at axed time--and an agent bidding in Amazon.com’s moretraditional English auction. The only rule different betweenthe two sites is the parameterthat controls the timingof the clearing event. However,that difference induces very dissimilar bidding strategies. In particular, ebay’s rules promotea strategy equivalentto that of the rst-price, sealed-bidauction: estimate the secondhighest bid and bid slightly aboveit. Amazon’s version of the auction induces the standard English type bidding behavior--keep bidding higher than the other buyersuntil either youwinor youare unwillingto pay more. Becausethere are potentially millions of waysto combine the auction parameters, we cannot makemuchprogress towardexible agents by analyzing individual auction types. Moreover,whenmultiple auctions are considered, the timing of the events and relationships betweenthe goodscan interact in complexways. Thus, we advocate automatic exploration of the space of strategies. Toaccomplishthis, we need a way to transform the auction rules into admissible actions. Oneapproachwhichwe are exploringis the use of declarative languagesfor reasoning about actions. The following description is meantto be suggestive of our approach.Consider a rule that de nes the action of placing a bid that is c morethan the current price quotein auction i: bid(i, quote(i, s) + e) ~ winning(i,s) A~elosed(i, s). Thisrule can be interpreted as: bid in auctioni for a price e morethan the current price quote if auction i is not closed andthe agent is currently winningzero units in i. This schemedependsuponus storing quite a bit of informationin the state o’. In particular, we needto store winning, closed, and quote for each auction. However, this informationis easily derived from the auction’s state transition function. Theaboverule is applicable only to single unit auctions. To develop a rule set applicable to the examplefrom Section 1, we needto introduce quantity to both the state and Auction closed 1 truc 2 false 3 false winning 1 0 0 bid 1 @$10 1 @$7 1 @$6 quote 1 @$10 1@$9 1 @$8 2 @$7.5 each Table1: A state in the thee auction example. the action rules. Table 1 showsa state that mayoccur for the three auction examplefrom Section 1. Thequote for the third auction displays prices that dependuponthe numberof units requested. The modied action rule is bid(i, quots(i,s,q) +e,q) ,in ning(i,s) < q A’=elosed(i,s). Thisrule states that an agent can place a bid in auction i for q units for a dollar value quote(i, s, q) + e if the auction is not closed and the agent is winningless than q units. Assumingthe state in Table 1 and that the domainof i is {1,2,3} and the domainof q is {1, 2}, the following conclusions can be drawn: bid(2, $9 + e, 1) bid(a, $8 + c, 1) bid(3, $15 + e, 2) Notice that this formulation permits the action bid(a,$15 + c,2) even though the agent has already wonan item in the rst auction. In this step weare computing all possibleactions. It is the task of the decisionprocess to evaluatethe utility of these actions. The above presentation is suggestive--a completeaction speci cation will require null actions and wait actions, and a variety of other time-relatedreasoning. 3 Decision Representations Wesee a natural connection betweendecision representations and the market model. Whenthe agent constructs a modelof the other individuals in the market, it can use game theory to select a strategy. However,whenthe market is modeledas a reactive black box, then MarkovDecisionProcesses are an appropriatetool. 3.1 Game-Theoretic Models Gametheory (Fudenberg &Tirole 1996), the fundamental tool for analysis in microeconomic settings, is applicableto decision makingin fragmentedmarkets. A particular fragmentedmarketscenario can be modeledas a strategic form gameusing standard notation. In gametheory terminology, the sequencesof bids are the agent’s strategies, and the sequences of bids by the other agents (or marketresponses) de ne the strategies of the opponent(s).Thus, the decision problem reduces to a problem of nding a Hash or BayesNashequilibrium. However,the strategic form is a cumber- 138 Nolicense: PDFproducedby PStill (c) F. Siegert - http://www.this.nel/~frank/pstill.html someview of the decision problem. It is more natural to view the analysis in an extensive form game. GameConstruction Let us continue with the scenario presented in Section 1. The agent is interested in buying two items for no more than $5 each, and has three auctions in whichit can participate. Supposethat the current price quotes from each auction are PA= $3, PB = $4, and Pc = $2 for one unit and $3 for two units. Supposethat the only admissiblebids in each auction are exactly $1 over the current price quote. If we viewthe agent’s problemas an extensive form game, the situation describedaboveis one branchon the gametree. This branch, and the actions associated with it are shownin Figure 3. The rectangle represents the gamestate, while the roundedrectangles indicate the agent’s possible actions. The actions are composedof bids placed in one or more of the three auctions (i.e. bc denotes placing a bid in auction C for the designatedvalue). Eachaction leads to a newgame state, whichin turn presents the agent with a newchoice of actions. In somecases, the agent maynot be able to precisely predict the outcomeof its actions. This can occur whenthe agent does not have complete information about the other agent’s actions in the auction, and thus cannotprecisely predict the effect that its bid will have on the auction state. Thesesituations can be modeledusing the notation for imperfect information in extensive form games(Fudenberg Tirole 1996). Strategy Selection The purpose of constructing a game-theoretic representation of the agent’s decision problem is to enable the agent to select a strategy. Considerable work has been done on computational game theory (Koller, Megiddo, & von Stengel 1994; 1996; McKelvey & McLennan 1996), and general purpose systems, such as GAMBIT (http://masada.hss.caltech.edu/~ambit/Gambit.html) and GALA (http://robot ics.stan ford.edu/koller/papers/gala.html), havebeen built for solving games. However,gametheoretic representations quickly become intractable as the numberof strategies and numberof agents increase. It maybe possibleto exploit regularities in the auction gamesto improvethe performanceof the general algorithms. For instance, in somesituations the agent’s decision problemmaybe effectively treated hierarchically--at a high level we mayrepresent the agent’s action in an ascending auction simply as the maximum bid it will place, and in a subplanenumeratingall of the intermediatebids. Similarly, whenthe valuesof bids are not restricted (as in the example), we need methodsto represent and reason with action ranges (eg. bid between$4 and $5) (cf. (Boutilier, Goldszmidt, Sabata 1999a)). 3.2 Markov Decision Models Wealso anticipate that somemodelsof the marketbehavior, particularly those that modelthe marketresponseas a purely reactive entity, will be amendableto decision making using MarkovDecision Processes (MDPs)(Boutilier, / PA = $3 PB = $4 Pc = $2 ~:)roneunJt $3 each ]br ~ o Figure 3: A portion of the gametree for an agent with a choiceof biddingin three auctions withdifferent rules. Dean, & Hanks 1999). In the simplest case, the market function generates a single marketaction, 7- and the state transition function is as described in Section 2.4. Thegeneral modelpresentedin this paper admits various sources of imperfectinformation,such as probability distributions over the agents’ preferences, af iiated common values, or simply the fact that the agent’s observationsof price quotes are an indirect measureof the other agent’sbids. In suchcases, the marketresponsefunction returns a set of possiblemarketactions whichin turn result in a set of possiblesuccessorstates. Thesesituations can be modeledusing Partially Observable MarkovDecision Processes (POMDPs). The agent frameworkwhich we are building will be designed to be modularso that we can implementboth of the above decision representations. Whenthe agent has several potential decisionrepresentationsavailable, it needsto performmetareasoningto select one to use for a particular probleminstance. This suggests another avenueof research, namely,the automaticselection of decision representation based on dynamicassessmentof problemfeatures. 4 Conclusion In this article we have advocatedan approachto building trading agents that stresses exibility. Wehave identi ed three inputs whichare central to a trading agent: user preferences, auction rules, and a marketmodel.Theseinputs feed the strategy generationand the strategy evaluation engines. Gametheory and decision theory, whencombinedwith an appropriate marketmodel,can be used for the task of strategy selection. Thearchitecture can handle a widevariety of auction rules and agent preferences. Weare currently building a trading agent in the spirit of the one describedhere. acknowledgments I would like to thank Stuart Feldman, AmyGreenwald, David Parkes, MikeWellman,and Weili Zhu for valuable conversationsregarding this work. 139 No license: PDFproduced by PStill (e) F. Siegert - http://www.this.net/~frank/pstill.html References Boutilier, C.; Dean, T.; and Hanks, S. 1999. Decisiontheoretic planning: Structural assumptionsand computational leverage. Journalof Arti cial Intelligence Research 11:1-94. Boutilier, C.; Goldszmidt,M.; and Sabata, B. 1999a. Continuous value function approximationfor sequential bidding policies. In Fifteenth Conferenceon Uncertaintyin Arti cial Intelligence. Boutilier, C.; Goldszmidt,M.; and Sabata, B. 1999b. Sequential auctions for the allocation of resources with complementarities. In Sixteenth InternationalJoint Conference on Arti cial Intelligence. Demange,G.; Gale, D.; and Sotomayor, M. 1986. Multiitem auctions. Journal of Political Economy 94:863-72. DeMartini, C.; Kwasnica, A. M.; Ledyard, J. O.; and Porter, D. 1998. A new and improved design for multiobject iterative auctions.Technicalreport, CaliforniaInstitute of Technology. Fudenberg, D., and Tirole, J. 1996. GameTheory. MIT Press. Gale, I. 1990. A multiple-objectauction with superadditive values. EconomicsLetters 34:323-28. Hu, J., and Wellman,M. P. 1998. Online learning about other agents in a dynamicmultiagent system. In Second International Conference on AutonomousAgents (Agents 98), 239-246. Koller, D.; Megiddo,N.; and von Stengel, B. 1994. Fast algorithmsfor nding randomizedstrategies in gametrees. In 26th ACMSymposium on the Theory of Computing, 750-759. Koller, D.; Megiddo,N.; and von Stengel, B. 1996. Efcient computationof equilibria for extensive two-person games. Gamesand EconomicBehavior 14(2):247-259. Lehmann,D.; O’Callaghan, L. I.; and Shoham,Y. 1999. Truth revelation in rapid, approximatelyef cient combina- torial auctions. In ACMConference on Electronic Commerce, 96-102. McKelvey,R. D., and McLennan,A. 1996. Computation of equilibrium in nite games. In Handbookof Computational Economics,volume1. Amsterdam:Elsevier Science, B.V. Park, S. 1999. The P-Strategy: An Adaptive Agent Bidding Strategy Based on Stochastic Modelingfor Continuous DoubleAuctions. Ph.D. Dissertation, University of Michigan. Parkes, D. C. 1999. /Bundle: Anefcient ascending price bundle auction. First ACMConferenceon Electronic Commerce 148-157. Rodr’guez,J.; Noriega,P.; Sierra, C.; and Padget,J. 1997. FM96.5a java-based electronic auction house. In Second International Conferenceon the Practical Application of lntelligent Agents and Multi-Agent Technology(PAAM ’97). Rust, J.; Miller, J. H.; and Palmer, R. 1993. Behaviorof Trading Automatain a ComputerizedDoubleAuction Market. Reading, MA:Addison-Wesley Publishing. chapter 6, 155-98. Wellman, M. P.; Walsh, W. E.; Wurman, P. R.; and MacKie-Mason,J. K. to appear. Auction protocols for decentralized scheduling. Gamesand EconomicBehavior. Wurman,P. R., and Wellman, M. P. 2000. AkBA:A progressive, anonymous-price combinatorial auction. In Second ACMConference on Electronic Commerce, 21-29. Wurman,P. R.; Weilman,M. P.; and Walsh, W. E. 1998. The MichiganInternet AuctionBot:A con gurable auction server for humanand software agents. In SecondInternational Conferenceon AutonomousAgents, 301-308. Wurman,P. R.; Wellman,M. P.; and Walsh, W.E. to appear. A parametrizationof the auction design space. Games and EconomicBehavior. 140 Nolicense: PDFproducedby PStill (c) F. Siegert - http://www.this.net/~ffank/pstill.html