Building coalitions involving agents and humans

Abstract

Agent-based participatory simulations are laboratory experiments designed like agent-based simulations and where humans access the simulation as software agents.This paper describes the outcomes of six experiments lasting up to two hours each, where human players took part in an iterated game derived from the El Farol bar problem. Agents decide synchronously to go to the bar or to stay home and the benefit depends on the bar attendance, with a threshold effect: it is better to stay home if more than 60% of the agents go. Contrasting with the original version of this problem, we allowed agents, and therefore humans, to communicate before they took their decision.The first two experiments allowed us to train participants and to introduce the notion of teams. Teams represented coalitions within the game and positively affected scoring, but they were not part of an obvious solution to the problem and they did not enforce cooperative behavior in the game.Drawing from these experiments, we designed autonomous agents reproducing strategies of the participants. These agents took part in the last four participatory experiments and we observed the formation of coalitions between agents, between humans and between agents and humans.