Evolutionary dynamics and optimal learning under bounded rationality

Individuals repeatedly face a sequential decision-making problem (or multi-decision bandit) in which the basic structure of which decision is faced under what circumstances is known whereas underlying payoff distributions are unknown. They have minimal memory and update their strategy by observing previous play (and not strategy) of someone else. We select behavioral rules that increase average payoffs as often as possible in a large population where all use the same rule. Here imitation generalizes to a pasting procedure. When decisions within the multi-decision bandit are unrelated, individuals eventually learn the efficient strategy but the underlying dynamic is not monotone. However, when choices influence which decisions are subsequently faced in the multi-decision bandit, play may not be efficient in the long run as it approaches a Nash equilibrium of the agent normal form. Copyright (C) 1998 IFAC.