Many is More, But Not Too Many: Dimensions of Cooperation of Agents with and without Predictive Capabilities

This paper examines the tradeoffs between agents that can predict (and, therefore, take into account) other agent’s actions and agents that act on their own without taking other agents’ actions into account. A simple prediction mechanism allows agents to make reasonably accurate guesses about other agents’ future actions, thereby making better decisions about their own actions. Our experimental findings from a multiagent object collection task suggest that not only do predictive agents perform better than non-predictive agents, but there are circumstances in which simple reactive prediction mechanisms perform as well as complex deliberative prediction mechanisms.