Applying Evolutionary Game Theory

to Auction Mechanism Design

Andrew Byde
Intelligent Enterprise Technology Laboratory
HP Laboratories Bristol
HPL-2002-321
November 27th , 2002*

E-mail: Andrew_Byde@[Link]

auctions, In this paper we describe an evolution-based method for evaluating

e-commerce, auction mechanisms, and apply it to a space of mechanisms
game theory, including the standard first- and second-price sealed bid auctions.
evolution, We replicate results known already in the Auction Theory literature
genetic regarding the suitability of different mechanisms for different
algorithms, bidder environments, and extend the literature by establishing the
multi-agent superiority of novel mechanisms over standard mechanisms, for
systems commonly occurring scenarios. Thus this paper simultaneously
extends Auction Theory, and provides a systematic method for
further such extensions.

* Internal Accession Date Only Approved for External Publication

 Copyright Hewlett-Packard Company 2002
 Applying Evolutionary Game Theory to Auction
Mechanism Design

Andrew Byde
Hewlett-Packard Laboratories,
Filton Road, Stoke Gifford,
Bristol BS34 8QZ, UK.
Andrew Byde@[Link]

ABSTRACT the space of potential market designs increases markedly,

In this paper we describe an evolution-based method for since mechanisms that might seem “non-sensical” or diffi-
evaluating auction mechanisms, and apply it to a space of cult to interpret for humans can be considered.
mechanisms including the standard first- and second-price
sealed bid auctions. We replicate results known already in Recently, a few papers have begun to address the confluence
the Auction Theory literature regarding the suitability of of these ideas, taking inspiration from the Auction Theory
different mechanisms for different bidder environments, and work on mechanism design, extending it into new design
extend the literature by establishing the superiority of novel spaces that might have been infeasible before, and adding a
mechanisms over standard mechanisms, for commonly oc- degree of automation to the design process: for example, in
curring scenarios. Thus this paper simultaneously extends [8] and [9], Cliff describes an application of Genetic Algo-
Auction Theory, and provides a systematic method for fur- rithms [13] to the choice of a continuous parameter Qs gov-
ther such extensions. erning the probability that a seller will be chosen to shout
in a given round of a stylized continuous double auction.
Keywords
Auctions, Economics, e-commerce This paper continues the of such work, examining a space
of auction mechanisms that includes the standard first- and
second-price auctions, using GAs applied to a multi-agent
1. INTRODUCTION system to evolve good players for each mechanism under
Auctions are an important class of mechanisms for resolving consideration. We find that under several classes of non-
multi-agent allocation problems of various types. There ex- pathological conditions (e.g. bidders are risk-averse, and
ists a substantial body of work (see, e.g. [14] for a review) are unaware of how many players they will face in a given
regarding the theory underlying auctions, most of which fo- auction), there exist exotic sealed bid mechanisms which
cuses on the problem of how to design them so as to achieve are expected to return significantly higher revenue to the
some desired outcome for the auctioneer. In situations where auctioneer than either the first- or second-price sealed bid
the auctioneer plays the role of seller, this outcome is often mechanisms. See Section 4 for more details.
revenue maximization, and many results of a qualitative na-
ture are known regarding the suitability of different mecha- The paper is laid out as follows: in the next two sections
nisms under different assumptions on the economic scenario we discuss the methods used, introducing relevant Game
under consideration. Theory concepts as needed. In Section 4 we describe the
results of our experiments. In Section 5 we describe in more
In parallel with this work, researchers have begun inves- detail the relationship between these results and others in
tigating how to design autonomous agents capable of par- the literature, and we conclude in Section 6.
ticipating in auctions ([18], [1], [4], [6], [7], etc.). Often
such study is motivated by the possibility that suitable au- 2. AUCTION THEORY
tonomous agents will be superior to humans in making (pos-
sibly quite complex) economic decisions, and indeed Das et
2.1 Terms and Notation
al. report human experiments to substantiate this possibil- In this paper we study sealed-bid auctions, in which a good
ity [11]. When the agents acting in markets are non-human, is put up for sale, and each potential buyer submits a bid
to the auctioneer; the auctioneer chooses a winner, and al-
locates payments to each agent. In most variants of this
type of auction, the good is awarded to the buyer who sub-
mits the highest bid, and only the winner pays. In a first-
price auction, the winner’s payment is equal to her bid; in
a second-price auction, the winner’s payment is equal to
the second highest bid.

In order to analyse how bidders might be expected to behave

in such auctions, we need to specify how they are motivated,

1
i.e. what is the good worth to each agent. We use a model and if (concerning the mechanism),
in which agents are only interested in their own awards and
payments, and there is some intrinsic monetary “value” v
• in equilibrium, the good always goes to the bidder with
associated to the good. All outcomes can therefore be rep-
the highest signal,
resented by a single number, the monetary gain the agent
makes. This is v − x for a win with payment x, and −x for • any bidder whose signal is the lowest possible expects
a non-win with the same payment. The risk preferences of to make nothing,
agents are differentiated by use of a von Neumann – Mor-
genstern utility function u, so that an agent strictly prefers
then the expected revenue to the seller is the same, inde-
a selection of possible outcomes xi with corresponding prob-
pendent of the mechanism.
abilities pi , over a second selection of possible outcomes yj
with corresponding probabilities qj , if and only if
X X This rather surprising result means that, subject to these
pi u(xi ) > qj u(xj ). (1) hypotheses, it doesn’t matter what type of auction a seller
i j runs, he should expect to make the same amount of money
whatever the mechanism. But of course there are many
In this representation, assuming twice-differentiability of u, different auction mechanisms in use, of extremely variable
an agent for which u00 (x) = 0 is known as risk-neutral; if type, because at least one of the hypotheses on which the
u00 (x) < 0, the agent is risk-averse, and if u00 (x) > 0, the Revenue Equivalence Theorem rests is often violated. It is
agent is known as risk-seeking (see Figure 1). known, for example, that most people are not risk-neutral1 ,
and in the case when the bidders are risk-averse, it makes
1.75 more sense for a seller to run a first-price auction.
Utility

1.5 Risk seeking

A method commonly used to establish an ordering on auc-
1.25
tions for different types of buyer, is to treat the problem as a
non-cooperative game, and solve for the game’s Nash equi-
1 librium. The difficulty with doing this is that the equations
used to define such an equilibrium might well be intractible.
0.75 Risk neutral In this paper we pursue an alternative method for determin-
ing an ordering: we simulate a population of buyers, and
0.5
Risk averse
play the game many times with random selections from this
0.25
population. The resulting averaged returns for the seller are
estimates of the true expected return. As such, the results
0.2 0.4 0.6
Surplus
0.8 1
they give are not as satisfactory as those derived from Game
Theory.

Figure 1: Typical utility functions for a risk- seek- 3. METHODS

ing, neutral and averse agent. The basic methodology pursued was to instantiate a group
of agents according to various environmental and agent-
The value of a good to an agent can be independent of the
preference parameters, and let them compete in a specified
value of the good to other agents, or it can be derived from
auction according to specified strategies, logging the utility
information about how other agents value the good. In the
extracted by each agent as a result. Since many of the envi-
former case we say that the agent has a private value for
ronmental parameters require some degree of randomization
the good, and in the latter case that there is some common
in the agent instantiation (e.g. randomized private value for
value component. We treat both these cases by postulat-
the good), this procedure was repeated a large number of
ing that each bidder receives a “signal”, and that the value
times, so as to generate an estimate of the expected utility
of the good to the agent is some specified function of all
to an agent of using a given strategy in a given context.
the agents’ signals. Since a bidder only necessarily knows
her own signal, her decision problem may in general involve
guess work about the worth of the good. 3.1 Context Parameterization: Environment,
Preferences and Mechanism
2.2 Revenue for the Auctioneer We chose to investigate a space of mechanisms very similar
In this paper we study mechanisms from the point of view of to the first- and second-price sealed bid auctions specified
the amount of money they are expected to make the seller. earlier.
Perhaps the most important result in this area is the Rev-
enue Equivalence Theorem, which states that if (con- Definition 1. Let w = (w1 , . . . , wn ) be a vector of n real
cerning the environment) numbers. A w-price auction is a sealed bid auction in which
the highest bidder wins the good, and pays
• there is a fixed number of bidders, known to everyone, PN
j=1 wj bidj
• all agents are risk-neutral, PN (2)
j=1 wj
• all bidders’ signals are picked from a common, known 1
Most people tend to act in a risk-averse manner in their
distribution, daily lives.

2
where N is the minimum of n and the number of bidders, for all players, was
and bid1 , bid2 , . . . are the bids, ordered highest to lowest. !
X 
uα d tj /n + (1 − d)vi − b . (5)
j
In this paper we examine a one-dimensional sub-space of w-
price auctions, namely those of type w = (1−w2 , w2 ). In this
parameterization, w2 = 0 is a standard first price auction, 3.2 Strategy Optimization
w2 = 1 a standard second-price auction, and all other values The above variables w2 , c, d, α specify the context in which
of w2 correspond to non-standard auction types that have the agents have to act, but not how they should act in that
not previously been studied. context. The most challenging piece of analysis in Mecha-
nism Design is always figuring out how a bidder is likely to
The space of agent preferences and environmental variables behave. The standard Game Theory approach is to enumer-
which we explored was motivated by examining exceptions ate all strategies that an agent might pursue, and determine
to the Revenue Equivalence Theorem; we allowed variable a strategy from which deviation is not rational, i.e. which
group size, variable risk preference, and correlated (non- is expected-utility maximizing given that the other agents’
independent) bidders’ signals. In addition, we allowed the behaviour is fixed.
degree of commonality in values to be altered.
As mentioned before, this process is often impossible, either
because of the intractability of the strategy space, or be-
1. The number of agents in each trial was either an arbi- cause the equations which need to be solved to determine a
trary fixed number, or was chosen with uniform prob- deviation-proof strategy are too complex.
ability from a set of consecutive integers bigger than
2. In most experiments the fixed number was chosen In this paper we take an empirical approach to finding good
to be 6, and the range {2, 3, 4, 5, 6}. strategies, whereby each agent in a population of bidders
is equipped with a bidding function which can be modified
2. The signals (t1 , . . . , tn ) of a group (a1 , . . . , an ) of bid- through evolution to adapt to the necessities of the game. As
ders were chosen to be a weighted sum of a shared the agents play the game, successful strategies are bred pref-
random signal and a sequence of independent random erentially, and thus the entire population improves. There
signals, with each such signal coming from a uniform is constant pressure to improve, because if an agent’s devia-
distribution on [0, 1]2 . Thus independent variables S, tion from the norm gives it a slightly higher expected utility,
X1 , . . . , Xn were generated, and the signal ti for agent then it will be slightly more likely to breed than average, and
ai was chosen to be c S + (1 − c)Xi , where c ∈ [0, 1] pa- so its genes will be preferentially reproduced into the next
rameterizes the degree of correlation between agents’ generation.
signals.
The main drawback to this approach is that it can neither be
3. The calculation of the utility extracted from winning guaranteed that the population will evolve a good strategy
the good depends on two properties of the agents in- within a reasonable period of time, nor that the solution on
volved: their risk preferences, and the degree of com- which the population eventually coverges is a global rather
monality in value. For risk, we chose to use Constant than local optimum. Thus we gain formal simplicity at the
Absolute Risk utility functions: cost of computation. We run the entire process of evolution
8 many times independently, and reduce the effect of mutation
< 1 (eαx − 1) if α 6= 0, as time goes by, so as to encourage convergence.
uα (x) = α (4)
: x if α = 0,
The link between genomes and bidding function was as fol-
α is zero for risk-neutral agents, negative for risk-averse lows: A gene consisted of a sequence (g1 , g2 , . . . , gk ) of real-
agents and positive for risk-seeking agents. Figure 1 valued “control points” assigned to evenly spaced input sig-
plots these functions for α = −1, 0, 1. nals (0, 1/k, . . . , 1). The bid output for an input signal t was
generated by interpolating between the control points:
To model common value, we assumed that the mone-
tary value
P to
agent ai of winning the good was given
by d · j vj /n + (1 − d)vi , where d is a parameter 8
controlling the degree of common value, with d = 0 < g1 if t < 0
representing purely private values, and d = 1 purely bid(t) = gl + (k.t − l)(gl+1 − gl ) if t ∈ [l/k, (l + 1)/k)
:
common values. Thus the utility reward to agent ai gk+1 if t ≥ 1
of winning the good at bid b, conditional on signals tj (6)
2
We also performed experiments with the family of distri- This representation was chosen over others (e.g. power series
butions representation, GP etc.) because it combines useful features
  of the domain, while placing very few restrictions on the
m−1
Bm,k (x) = m · xm−k (1 − x)k k − 1 (3) space of all such functions. Specifically,

(for k = 0, . . . , m). The results generated with these distri-

butions were not qualitatively different from those generated 1. Stability: These functions are stable under small ran-
for the uniform distribution. dom mutations: changing the data ci does not make a

3
 huge difference to the output values generated by the are very unlikely to win the good (and hence have nega-
bid function. tive surplus), whereas they are much more likely to decrease
the winner’s surplus, and hence increase their own relative
2. Locality: A change in a value gl has no effect on the fitness.
function for signals above (l + 1)/n or below (l − 1)/n,
so each g value, or sequence of g values, represents a Thus this process finds good players at the repeated com-
partial solution for a certain input range. petitive game, not at the one-shot game. It is hoped (and
we shall demonstrate, in some cases) that this effect is very
In addition, the functions generated are guaranteed to be small, so that conclusions about the one shot game can still
continuous. be made; it is worth noting that in real auctions with real
players (humans or corporations), exogenous effects such as
These data gl were then mutated and recombined according these are commonplace.
to a standard Genetic Algorithm, where the fitness of a given
genome was determined relative to other genomes by par- 4. RESULTS
ticipation in a sequence of randomized games. Specifically,
Shown below (Figure 2) is a graph showing the genes of
the evaluation of a population of genomes was according to
the best individual in a population of 360, plotted against
the following algorithm:
generation number, for an example evolutionary run for risk-
For each of a large number of iterations { neutral agents with independent private values, competing
while (not all agents have played in this round) { in a second price auction. In this context, the optimal bid-
select some as-yet-unplayed agents to play a game
ding strategy is to bid one’s signal, so given that k = 5
generate random signals for the agents
get bids for each agent, according to their genome (i.e. the genome consists of 5 control values), the expected-
select a winner and determine payments utility-maximizing genome is (0, 0.25, 0.5, 0.75, 1.0). As can
accumulate the corresponding utility rewards be seen, the best individual is initially far from perfect, and
} varies greatly over the first few generations, since mutation
} is (relatively) high, and the population very diverse. As
time wears on, however, the population discovers a bidding
An agent’s fitness was equal to its accumulated utility from strategy that is close to optimal: the final population’s best
all the games. This process is modular with respect to the individual’s bid is always within 5% of its optimal value.
contextual parameters specified in Section 3.1.
1.25

The Genetic algorithm was simply

generate a population of N random genomes 1
for each generation {
assess the fitness of every individual by
playing a large number of games as above 0.75

rank the genomes by this fitness measure

repeat N times { 0.5
select two genomes from the old population,
favouring highly ranked genomes
select a random point in the genome, and 0.25
combine the first half of one with the
second half of the other
with pre-selected probability, mutate each 0

of the genes by an amount picked from a 0 20 40 60 80 100 120 140 160 180 200

pre-selected distribution3
-0.25
place this new genome into the new population
}
} Figure 2: Genome for best individual in an example
evolutionary run, plotted as a function of generation
Thus we did not necessarily preserve the fittest individual [Link] population is initialized to random ra-
from each generation. Notice that the fitness function is tional strategies (i.e. agents in the initial population
stochastic, so genomes can gain unfair advantage from being cannot lose money, initially).
lucky (in the selection of their signals, for example). Notice
also that the fitness is measured relative to other agents. We first verified the method by calculating revenue land-
This means that the most successful agent strategy is not scapes for situations where the ordering of first and sec-
necessarily that which gives greatest expected return, since ond price auctions is qualitatively known. When players are
it may (for example) be incentive compatible in such an en- risk-neutral, signals independent, and the number of play-
vironment to deliberately disadvantage oneself if in doing so ers fixed, the Revenue Equivalence Theorem says that all
ones opponents are even more disadvantaged. An example our mechanisms will generate the same expected revenue.
of this is that agents with very low signals are (in most envi- Figure 3 plots expected seller revenue against w2 (in black).
ronments) incented to bid higher than their valuation: they These results were obtained by evolving a population of 360
3
The mutation probability was constant (at 0.8), and the agents for 200 generations, 200 times, and taking the aver-
maximal size of the mutation µ was reduced as time went age auctioneer revenue across these 200 trials. The revenue
on, being given, in generation G, by µ = µ0 (G0 /(G0 + G))λ , was always calculated on the basis of all agents using the
for fixed µ0 = 0.05, G0 = 20, and λ = 1.5 best individual strategy from generation 200. The two lines

4
 0.62
in grey, above and below the plotted curve of average rev-
enue, are plus and minus one standard deviation relative
to the average, and give an indication of the magnitude of
experimental uncertainty. 0.61

As can be seen, there is no experimentally significant dif-

ference in revenue between any of the mechanisms in the 0.6
risk-neutral independent private values case.
0.73
0.59

0.72
0.58
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.71 Figure 5: Average revenues for risk-neutral agents

in fixed-size groups with values that are 50% corre-
lated.
0.7

investigating regions of the environment space where there

are no clear cut results. For example, if buyers have partial
common values, and are risk averse, then either first or sec-
0.69
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 ond price could be optimal for the seller, depending on the
magnitude of the two effects.
Figure 3: Sample revenues for risk-neutral agents
0.62
operating in groups of fixed size 6, versus coefficient
w2 of second-highest bid in payment. Left-hand side
is first price auction, right-hand side is second-price
auction.
0.61
As mentioned in Section 2.2, when we modify the above by
having risk-averse buyers, the first price auction becomes
preferred. Figure 4 shows this effect.
0.6
0.77

0.76

0.75 0.59

0.74

0.73 0.58
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.72
Figure 6: Average revenues for risk-averse agents
0.71 (α = −10) in fixed-size groups with values that are
50% common.
0.7
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 6 shows the situation when bidders are risk-averse,
with parameter −10, and have a common value coefficient
Figure 4: Average revenues for risk-averse agents of 0.5. In this case, the first-price auction is clearly superior
operating in groups of fixed size 6. to the second-price auction. More surprisingly, a (0.3, 0.7)-
price auction is superior to both first- and second-price auc-
As Milgrom et al. demonstrate, when values are correlated4 , tions.
we expect that the second-price auction will give greater
revenue [15]. Figure 5 demonstrates this effect occurring. Figure 7 shows the same situation when the common value
coefficient is 0.9, and risk aversion is −15. In this case,
Much more interesting than confirming known results, is second-price is superior to first-price, and once again a w-
4
In fact [15] discuss not correlation but affiliation between price auction is superior to both. These auction forms, in
bidders’ signals. It can be shown that the joint signal dis- which the winner pays a weighted average of his own and
tributions we have chosen to use in this paper satisfy this the second player’s bid are not studied in the literature, but
stronger affiliation condition. in this common scenario, can be revenue maximizing, de-

5
 0.49 0.005

0.48 0.0045

0.47 0.004

0.46 0.0035

0.45 0.003
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 7: Average revenues for risk-averse agents Figure 9: Graph of expected utility earned by each
(α = −15) in fixed-size groups with values that are agent as a function of w1 for the equilibrium strategy
90% common. in the scenario described in Figure 7

pending on the nature of the agents playing the game. The

optimality of non-standard auctions in this risk-averse, par-
5. RELATED WORK
Auction Theory is a mature field, with a substantial lit-
tial common-value setting persists if the number of agents
erature. We shall not attempt an exhaustive review here;
is variable, as is seen in Figure 8.
interested readers are referred to [14] for an overview.
0.43
The use of agents to investigate economic phenomena via
simulation is a much newer field, known, broadly, as Agent-
based Computational Economics [22], a sub-field of which is
concerned with designing agents that can operate in on-line
0.42
auctions or negotiations. See, for example [1], [6] with re-
spect to 1-1 negotiation; [7], [19], [12], [11], [21] with respect
to continuous double auctions; [17], [18], [5] with respect to
sequences of English Auctions and [4], [3] with respect to
0.41 sequences of sealed bid auctions.

When it comes to investigating novel auction types automat-

ically, or semi-automatically, the citations are much thinner
0.4
on the ground. A general discussion of automated mecha-
nism design appears in [10], which deals with issues of com-
putational complexity, but does not address any practical
implementation details. The work of Cliff [8] is the first to
provide a complete system for automated mechanism design.
0.39
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Cliff addresses the case of a continuous open-cry auction, us-
ing a Genetic Algorithm to adjust both the parameters of
Figure 8: Average revenues for risk-averse agents the bidding agents he uses, and the mechanism parameter,
(α = −15) with values that are 90% common, when which in this case is the probability Qs that in any given
group size is random in the range [2, 6]. round, a seller will be chosen at random to make an of-
fer. Besides being based on the continuous double auction,
Cliff’s work differs from ours in two significant ways. Firstly,
In both of these scenarios, the graph of utility versus w1
the space of agent strategies explored is necessarily very re-
is flat (see Figure 9): the agents themselves are indifferent
strictive5 , whereas the strategy space our GA explores con-
as to which auction they participate in. Thus selecting a
tains close approximations to all continuous bidding strate-
revenue-maximizing value of w1 need not antagonize bid-
gies. Secondly, although Cliff’s choice of mechanism space
ders, however, as Bergman et al. show in [2], we should
expect the real dynamics of auction choice on the part of 5
This problem difficult to address in the continuous open-
bidders to be affected by more than just expected revenue: cry auction, because such auctions have an intractably com-
the variance of payments is crucial also. Clearly more work plicated strategy space: an agent will typically have many
is needed to understand population dynamics in this new opportunities to act, for each of which the information space
environment. is the set of all previous actions by all agents.

6
was inspired by the experimental design used by Smith in [7] D. Cliff and J. Bruten. Less than human: Simple
[20], the Continuous Double Auction as it is used in such adaptive trading agents for CDA markets. In Proc. of
real-world institutions as the New York stock exchange is the 1998 Symposium on Computation in Economics,
quite different – using order queues and bid improvement Finance, and Engineering: Economic Systems, 1998.
rules, for example; our cases w2 = 0, 1 are faithful interpre-
[8] Dave Cliff. Evolution of market mechanism through a
tations of first- and second-price sealed bid auctions, which
continuous space of auction-types. Technical Report
are used in the world on a daily basis.
HPL-2001-326, HP Labs, 2001.
The work of Phelps et al. [16] provides another approach [9] Dave Cliff. Evolution of market mechanism through a
to modifications of the continuous double auction, in which continuous space of auction-types ii: Two-sided
the modification is to the clearing rule, via use of Genetic auction mechanisms evolve in response to market
Programming. shocks. Technical Report HPL-2002-128, HP Labs,
2002.
6. CONCLUSIONS [10] V. Conitzer and T. Sandholm. Complexity of
In this paper we have described an application of simulated mechanism design. In Proc. 18th Conference on
evolutionary game theory to a mechanism design problem. Uncertainty in Artificial Intelligence (UAI), August
We have demonstrated that this technique can be used to 2002.
explore a space of auction mechanisms, and by doing so
in a specific setting that involves faithful versions of real- [11] R. Das, J.E. Hanson, J.O. Kephart, and G. Tesauro.
world mechanisms, have established the superiority of non- Agent-human interactions in the continuous double
standard auction types in a variety of common environ- auction. In Proc. 17th Int. Joint Conf. on Artificial
ments. Intelligence, 2001.
[12] S. Gjerstad and J. Dickhaut. Price formation in
The advantages of such a method for exploring auction de- double auctions. Games and Economic Behaviour,
sign issues are clear: the agents discover good bidding strate- 22(1):1–29, 1998.
gies by evolution, without the need for complicated, possi-
bly intractable, and certainly fragile mathematical analy- [13] J. Holland. Adaption in Natural and Artificial
sis. In more complicated applications, the evolution process Systems. U. Michigan Press, 1975.
can implicitly take factors into consideration that might not [14] P. Klemperer. Auction theory: A guide to the
have occurred to analysts. Additionally, the mechanism is literature. Journal of Economic Surveys,
tested for revenue generation against a small neighborhood 13(3):227–286, July 1999.
of strategies, not just the Nash-equilibrium strategy. As a
result, its sensitivity to agents’ choice of strategy can be [15] M. P. Milgrom and R. J. Weber. A theory of auctions
determined. and competetive bidding. Econometrica, (50):1089 –
1122, 1982.
7. REFERENCES [16] Steve Phelps, Peter Mc Burnley, Simon Parsons, and
[1] P. Anthony, W. Hall, V. Dang, and N.R. Jennings. Elizabeth Sklar. Co-evolutionary auction mechanism
Autonomous agents for participating in multiple design. In Lecture Notes in AI, volume 2531. Springer
on-line auctions. In Proc. IJCAI Workshop on Verlag, 2002.
E-Business and the Intelligent Web, pages 54–64,
[17] C. Preist, C. Bartolini, and I. Philips. Algorithm
2001.
design for agents which participate in multiple
[2] A. Bergman and M. Tennenholtz. On the natural simultaneous auctions. In F. Dignum and U. Cortes,
selection of market choice. Autonomous Agents and editors, Agent Mediated Electronic Commerce III,
Multi-Agent Systems, 5(4):387 – 396, 2002. Lecture Notes in AI. Springer Verlag, September 2001.
[18] C. Preist, A. Byde, and C. Bartolini. Economic
[3] C. Boutilier, M. Goldszmidt, and B. Sabata. dynamics of agents in multiple auctions. In Proc. 5th
Continuous value function approximation for Int. Conf. on Autonomous Agents, 2001.
sequential bidding policies. In Proc. UAI ’99, 1999.
[19] C. Preist and M. van Tol. Adaptive agents in a
[4] C. Boutilier, M. Goldszmidt, and B. Sabata. persistent shout double auction. In Proc. 1st Int.
Sequential auctions for the allocation of resources with Conf. on the Internet, Computing and Economics.
complementarities. In Proc. IJCAI ’99, 1999. ACM Press, 1998.

[5] A. Byde. A dynamic programming model for [20] V. L. Smith. Experimental study of competetive
algorithm design in simultaneous auctions. In Proc. market behaviour. J. of Political Economy, 70:111 –
2nd Int. Workshop on Electronic Commerce 137, 1962.
(WELCOM-01), LNCS. Springer Verlag, 2001. [21] G. Tesauro and J. L. Bredin. Strategic sequential
bidding in auctions using dynamic programming. In
[6] A. Byde, C. Preist, and N. Jennings. Decision Proc. AAMAS02, pages 591 – 598, 2002.
procedures for multiple simultaneous auctions. In
Proc. 1st Int. Conf. Autonomous Agents and [22] Leigh Tesfatsion. Agent-based computational
Multi-Agent Systems, 2002. economics: Growing economies from the bottom up.