Technical Perspective: The Complexity of Computing Nash Equilibrium

Computer science and game theory go back to the same individual, John von Neumann, and both subjects deal with the mathematization of rational decision making. Yet, for many years they continued to work apart. Computer science concentrated mostly on issues of complexity; game theory mostly on issues of incentives.

But in the last dozen years we have seen a fusion of the two fields. Methodologies of each are used to solve problems in the other, and the concepts of each are incorporated to create better and more robust models in the other.

Nash equilibrium, introduced in the 1950s, is the main concept used in the analysis of strategic games. The equilibrium is simple to describe, logically appealing, powerful in making predictions, and its existence was brilliantly demonstrated by John Nash.

Despite its importance, research in the possibility of computing Nash equilibrium for games with many strategies only began in the late 1980s with studies by Gilboa and Zemel. This question was taken on full force by Papadimitriou and coauthors in a series of breakthrough papers. The research presented here is central to this literature, and (together with a follow-up paper by Chen and Deng for the case n=2) it offers a complete negative result. Computing a Nash equilibrium of an arbitrary n-person non-cooperative game with many individual strategies is at least as difficult as any problem that belongs to the class of PPAD-complete problems believed, and is considered too difficult for practical computations.

The message to users of Nash equilibrium may be devastating, as they ask: "If my computer cannot compute it, how can players in the market do it?" The negative results the authors report here even apply to algorithms where players can be centrally coordinated, and in the situations being modeled, the problem is actually even more intractable, since a solution must be found by a dispersed, uncoordinated set of players.

But as is often the case with impossibility results, this one leads to a large variety of follow-up questions dealing with the assumptions, the methodology, and the relevance.

The standard worst-case approach employed by the authors here assumes that for every proposed algorithm, one should consider the most difficult n-person game to compute, and study the computation time as the number of individual strategies becomes arbitrarily large.

Contrary to this approach, it is argued that a game designed to defeat an algorithm is not likely to be natural in applications. As a result, we have seen alternative approaches and conclusions to the question of computing a Nash equilibrium in games with many strategies.

First, motivated by economic, political, and other applications, there have been studies of computations of Nash equilibrium for restricted classes of games: games with anonymous opponents, games with graphical structures, games with continuous and/or convex payoff functions, and more. As it turns out, in many of these games researchers are able to establish positive results.

There are positive results on computing—with a high probability—an equilibrium of a many-strategies randomly generated game. This approach, however, must assume a probability distribution by which games are generated, and the results may depend on the distribution.

Finally, an old established approach is to ignore the asymptotic nature of the question, and simply find algorithms that work well in practice (similar to the simplex algorithm for linear programming working well, despite its asymptotic algorithmic inefficiency). Recent research leads to algorithms that are efficient in computing equilibria of rich classes of test games.

Back to Top

Back to Top