departure from the bulk of the equilibrium refinements literature in distinguishing
between contending strict Nash equilibria.
This ability is not an unqualified success. The limiting stationary distribution,
as the probability of a mutation goes to zero, may be a reasonable approximation
only of what occurs after long periods of time, with this waiting time becoming very
long when the probability of a mutation is quite small. In some cases, the implied
waiting times will be sufficiently long that our interest will center on shorter
horizons and the stationary distribution will be irrelevant. In other cases, the
stationary distribution may be more useful. Beginning with the work of Ellison
(1993), it has been recognized that waiting times may be significantly reduced if
there are spatial or “local” patterns to the interaction between agents. Young (1998)
discusses the evolution of social structures that may occur over sufficiently long
periods of time for the theory to be applicable. Much remains to be done, but it is
clear that evolutionary game theory has provided new tools to address a difficult
question.
Equilibrium Refinements
To see how evolutionary game theory does less than the equilibrium refinements
literature, we return to the cornerstone of the refinements literature: the
presumption that weakly dominated strategies should not be played. Consider the
game whose normal and extensive forms are shown in Figure 5. Binmore, Gale and
Samuelson (1995) interpret this as a simplified version of the ultimatum game.
Player 1 must propose an amount of a surplus of size 4 to offer to player 2 and can
choose either a high offer of 2 or a low offer of 1. A high offer is assumed to be
accepted, while a low offer may be either accepted (“Yes”) or rejected (“No”).
An equilibrium of this game is subgame perfect if player 1’s choice is a best
response to player 2’s choice and if player 2’s Yes/No decision would be a best
response in the event that player 1 chooses Low. Backward induction identifies the
only subgame perfect equilibrium of this game: player 2 accepts low offers, and as
a result, player 1 makes a low offer. There are other Nash equilibria in which player
1 makes a high offer and player 2 plays No with probability at least 1/3.
No is a dominated strategy for player 2. It can never earn a higher payoff than
Yes, and one would accordingly expect an evolutionary process to exert constant
pressure against No. Suppose, however, that a large fraction of the player-2 population
initially plays No. High will then produce a higher average payoff for player
1 than Low, and an evolutionary process will also exert pressure against Low. But
as fewer and fewer player 1s choose Low, the payoff disadvantage of No dissipates,
and hence, so does its evolutionary disadvantage. The result may be convergence to
an outcome in which player 1 offers High and a significant fraction of player 2s
would reject Low if offered. The dominated strategy No is thus not eliminated.
Binmore, Gale and Samuelson (1995) and Roth and Erev (1995) fill in the
details of this argument. However, it seems as if this argument relies too heavily on
the fact that player 2s’ choice of No is never tested if player 1s make high offers. We
expect the world to be a noisy place, certainly noisier than our simple models. It