Heuristics Recognition and Learning in Rock-paper-scissors Game: Experimental Study

Abstract:

The classical theory of learning in repeated games considers learning as a reaction to success or failure of a specific choice in previous rounds. However, in practice there can be other rules of learning: for example, people can spot specific patterns in opponents’ behavior and, thus, predict their future behavior. We study to what degree this type of learning is successful on the example of laboratory game of «rock-paper-scissors». Our participants — 70 students and schoolchildren from Moscow — played this game for 100 rounds against a computer algorithm, which was programmed to play optimally against a boundedly rational player. We show that participants successfully recognize regularities in computer opponent and are able to learn to optimally reply to such program. Success in recognition is directly proportional to program’s predictability (share of nonrandom moves by computer). Moreover, players are better at learning the pattern that allows them to win and are worse at learning from defeats. The results suggest that people can successfully use procedurally-rational strategies that are based on rule learning, and not only on reactions to past successes and failures.