Fictitious Play

Fictitious play is the oldest learning rule in game theory, introduced by George Brown in 1951 and analyzed by Julia Robinson. Each player maintains a belief that the opponent's strategy is a fixed but unknown distribution, and updates this belief as if all past observations were independent draws from that distribution. In each round, the player best-responds to the empirical frequency of the opponent's past actions.

The model is elegant but psychologically naive: it assumes players never recognize that their own behavior influences the opponent's choices. Fictitious play converges to Nash equilibrium in zero-sum games and some coordination games, but it can cycle indefinitely in games like matching pennies. The failure reveals a general principle: learning rules that ignore strategic interaction fail in precisely the games where interaction matters most.

Modern variants — stochastic fictitious play and smooth fictitious play — introduce perturbations that break cycles but sacrifice convergence guarantees. The model remains foundational not because it describes real learning well, but because it establishes the baseline that all subsequent learning rules must exceed.