Abstract. Experimentalists frequently claim that human subjects in the laboratory violate game-theoretic predictions. It is here argued that this claim is usually premature. The paper elaborates on this theme by way of raising some conceptual and methodological issues in connection with the very definition of a game and of players ’ preferences, in particular with respect to potential context dependence, interpersonal preference dependence, backward induction and incomplete information.

According to conventional wisdom, Nash equilibrium in a game “involves” common knowl-edge of the payoff functions, of the rationality of the players, and of the strategies played. The basis for this wisdom is explored, and it turns out that considerably weaker conditions suffice. First, note that if each player is rational and knows his own payoff function, and the strategy choices of the players are mutually known, then these choices form a Nash equilibrium. The other two results treat the mixed strategies of a player not as conscious randomization of that player, but as conjectures of the other players about what he will do. When n = 2, mutual knowledge of the payoff functions, of rationality, and of the conjectures yields Nash equilibrium. When n ≥ 3, mutual knowledge of the payoff functions and of rationality, and common knowl-edge of the conjectures yield Nash equilibrium when there is a common prior. Examples are provided showing these results to be sharp.

We present a formalization of the Recursive Modeling Method, which we have previously, somewhat informally, proposed as a method that autonomous artificial agents can use for intelligent coordination and communication with other agents. Our formalism is closely related to models proposed in the area of game theory, but contains new elements that lead to a different solution concept. The advantage of our solution method is that always yields the optimal solution, which is the rational action of the agent in a multi-agent environment, given the agent's state of knowledge and its preferences, and that it works in realistic cases when agents have only a finite amount of information about the agents they interact with. Introduction Since its initial conceptual development several years ago (Gmytrasiewicz, Durfee, & Wehe 1991a; 1991b), the Recursive Modeling Method (RMM) has provided a powerful decision-theoretic underpinning for coordination and communication decisionmaking, including dec...

This paper extends the framework of partially observable Markov decision processes (POMDPs) to multi-agent settings by incorporating the notion of agent models into the state space. Agents maintain beliefs over physical states of the environment and over models of other agents, and they use Bayesian update to maintain their beliefs over time. The solutions map belief states to actions. Models of other agents may include their belief states and are related to agent types considered in games of incomplete information. We express the agents ’ autonomy by postulating that their models are not directly manipulable or observable by other agents. We show that important properties of POMDPs, such as convergence of value iteration, the rate of convergence, and piece-wise linearity and convexity of the value functions carry over to our framework. Our approach complements a more traditional approach to interactive settings which uses Nash equilibria as a solution paradigm. We seek to avoid some of the drawbacks of equilibria which may be non-unique and are not able to capture off-equilibrium behaviors. We do so at the cost of having to represent, process and continually revise models of other agents. Since the agent’s beliefs may be arbitrarily nested the optimal solutions to decision making problems are only asymptotically computable. However, approximate belief updates and approximately optimal plans are computable. We illustrate our framework using a simple application domain, and we show examples of belief updates and value functions. 1.

Understanding knowledge is a fundamental issue in many disciplines. In computer science, knowledge arises not only in the obvious contexts (such as knowledgebased systems), but also in distributed systems (where the goal is to have each processor "know" something, as in agreement protocols). A general semantic model of knowledge is introduced, to allow reasoning about statements such as "He knows that I know whether or not she knows whether or not it is raining." This approach more naturally models a state of knowledge than previous proposals (including Kripke structures). Using this notion of model, a model theory for knowledge is developed. This theory enables one to interpret the notion of a "finite amount of information". A preliminary version of this paper appeared in Proc. 25th IEEE Symp. on Foundations of Computer Science, 1984, pp. 268--278. This version is essentially identical to the version that appears in Journal of the ACM 38:2, 1991, pp. 382--428. y Part of th...

We survey the recent work in AI on multi-agent reinforcement learning (that is, learning in stochastic games). We then argue that, while exciting, this work is flawed. The fundamental flaw is unclarity about the problem or problems being addressed. After tracing a representative sample of the recent literature, we identify four well-defined problems in multi-agent reinforcement learning, single out the problem that in our view is most suitable for AI, and make some remarks about how we believe progress is to be made on this problem.

In modelling competition among mechanism designers, it is necessary to specify the set of feasible mechanisms. These specifications are often borrowed from the optimal mechanism design literature and exclude mechanisms that are natural in a competitive environment; for example, mechanisms that depend on the mechanisms chosen by competitors. This paper constructs a set of mechanisms that is universal in that any specific model of the feasible set can be embedded in it. An equilibrium for a specific model is robust if and only if it is an equilibrium also for the universal set of mechanisms. A key to the construction is a language for describing mechanisms that is not tied to any preconceived notions of the nature of competition.

In their seminal paper, Mertens and Zamir (1985) proved the existence of a universal Harsanyi type space which consists of all possible types. Their method of proof depends crucially on topological assumptions. Whether such assumptions are essential to the existence of a universal space remained an open problem. We answer it here by proving that a universal type space does exist even when spaces are defined in pure measure theoretic terms. Heifetz and Samet (1996) showed that coherent hierarchies of beliefs, in the measure theoretic case, do not necessarily describe types. Therefore, the universal space here differs from all previously studied ones, in that it does not necessarily consist of all We study here the foundations of that part of the theory of games with incomplete information that deals with players ’ beliefs. We study it in the broadest and most natural setup, that of probability (or measure) theory without any topological notions, which have always been used for this purpose until now. We show that even under this general setup

The Common Prior Assumption (CPA) is central to the economics of information and the foundations of game theory. Recent contributions (Dekel and Gul, 1997, Gul, 1996, Lipman, 1995) have questioned its meaningfulness in situations of incomplete information where there is no ex ante stage and the primitives of the model are the individuals ’ belief hierarchies. We address this conceptual issue by providing characterizations of two local versions of the CPA which are in terms of the primitives and, therefore, do not involve a counterfactual and problematic ex ante stage. The characterizations involve three notions: Comprehensive Agreement, no error of beliefs and common belief in no error. Comprehensive Agreement is defined as the absence of “agreement to disagree ” about any aspect of beliefs; it is a generalization of Aumann’s (1976) notion of agreement. The entire analysis is carried out locally, that is, with reference to the “true state ” (which represents the actual profile of belief hierarchies) and does not rely on the Truth Axiom for individual beliefs. The results are also applied to the problem of generalizing the notion of Bayesian updating to single-person, intertemporal situations without perfect recall and without given information partitions. We are grateful to Bart Lipman and two referees for helpful and constructive comments. Seminar participants at Harvard, Penn, Princeton, USC and Yale provided useful comments. We also greatly benefited from discussions with participants at the SITE Workshop on the Epistemic Foundations of Game Theory (Stanford), in particular Steve Morris.

We present a distribution-free model of incomplete-information games, both with and without private information, in which the players use a robust optimization approach to contend with payoff uncertainty. Our “robust game” model relaxes the assumptions of Harsanyi’s Bayesian game model, and provides an alternative distribution-free equilibrium concept, which we call “robust-optimization equilibrium, ” to that of the ex post equilibrium. We prove that the robust-optimization equilibria of an incomplete-information game subsume the ex post equilibria of the game and are, unlike the latter, guaranteed to exist when the game is finite and has bounded payoff uncertainty set. For arbitrary robust finite games with bounded polyhedral payoff uncertainty sets, we show that we can compute a robust-optimization equilibrium by methods analogous to those for identifying a Nash equilibrium of a finite game with complete information. In addition, we present computational results.