Bounded-Rational Behavior by Neural Networks in Normal Form Games

This paper presents a neural network model developed to simulate the endogenous emergence of bounded-rational behavior in normal-form games. There exists an algorithm which, if learnt by a neural network, would enable it to perfectly select Nash equilibria in never before seen games. However, finding this algorithm is too complex a task for a biologically plausible network, and as such it will instead settle for converging to an approximation to Nash in a subset of games. We employ computer simulations to show that Nash equilibria are found approximately 60% of the times, and to characterize the behavioural heuristics acquired by the bounded-rational agent. Pure sum of payoffs dominance, and the best response to this strategy, get closest to predicting the networks behavior.