Learning under minimal information

An experiment on mutual fate control

Abstract

Reinforcement learning has proved quite successful in predicting subjects' adjustment behaviour in repeatedly played simple games. However, reinforcement learning does not predict convergence to the efficient cell in the minimal information game of mutual fate control, while earlier psychologist' experiments show some tendency to convergence. Our rivalling learning rule, a modification of win-stay lose-change, does predict convergence. We perform an experiment using modern economic methodology and compare these tow learning rules. Our results are unfavourable for both reinforcement learning as well as win-stay lose-change. The data rather support the view that subjects search by using patterns.