Neural Networks as a Learning Paradigm for General Normal Form Games

Abstract

This paper models the learning process of a population of randomly-rematched tabula rasa neural network agents playing randomly generated 3 × 3 normal form games of all strategic types. Evidence was found of the endogenous emergence of a similarity measure of games based on the number and types of Nash equilibria, and of heuristics that have been found effective in describing human behavior in experimental one-shot games. The neural network agents were found to approximate experimental human behavior very well across various dimensions such as convergence to Nash equilibria, equilibrium selection and adherence to principles of dominance and iterated dominance. This is corroborated by evidence from five studies of experimental one-shot games, as the Spearman correlation coefficients of the probability distribution over the neural networks’ and human subjects’ actions ranged from 0.49 to 0.89.