Algorithmic bounded rationality, optimality and noise

Abstract

A model of learning, adaptation and innovation is used to simulate the evolution of Moore machines (executing strategies) in the repeated Prisoner's Dilemma stage-game. In contrast to previous simulations that assumed perfect informational and implementation accuracy, the agents' machines are prone to two types of errors: (a) action-implementation errors, and (b) perception errors. The impact of bounded rationality on the agents' machines is examined under different error-levels. The computations indicate that the incorporation of bounded rationality is sufficient to alter the evolutionary structure of the agents' machines. In particular, the evolution of cooperative machines becomes less likely as the likelihood of errors increases.