Win-Stay, Lose-Shift. A General Learning Rule for Repeated Normal Form Games

Martin Posch and Werner Brannath - University of Wien

In this work we study a simple learning paradigm in the context of iterated normal form games. Following H. A. Simon's concept of satisficing we design players with a certain aspiration level. If their payoff is below this level, they change their action, otherwise they repeat it. Pavlov, the most simple win-stay, lose-shift strategy has been found to be a very robust strategy for the iterated Prisoners Dilemma that outperforms the celebrated Tit for Tat in the presence of noise. We study the efficiency of win-stay, lose-shift strategies for other games and in environments where many different games are played. A special emphasis is put on the impact of noise. We consider stochastic modifications of Pavlov that make errors, as well as strategies that average the received payoff over some rounds of the game before comparing it to their aspiration level. For high noise levels this averaging becomes favorable.

Scheduled for Session 3.3 Learning