Improving learning in robot teams through personality assignment

Adaptive robotic teams, acting as Multi Agent Systems (MAS) of concurrently learning agents, use reward or punishment as reinforcement while learning to take optimal actions. The utility of reinforcement learning in single agent systems has been demonstrated in the context of MAS, but agent interactions make learning actions that best benefit the entire team challenging. When rewards are local to each agent, they often do not learn to take altruistic actions. Small teams often learn cooperation when they earn global rewards from individual agents' actions. However, in large teams the global reward may not be attributable to a particular agent's actions, impairing their ability to learn effectively. This issue, commonly known as the Credit Assignment Problem for MAS, is addressed.;The first approach adopted assigns roles to each agent and loosely couples their rewards to achieve a blend of local and global reward systems. The reward system is applied to various combat scenarios. Developed algorithms are simulated in a tank combat environment, called Robocode, and a MAS consisting of a driver agent and gunner agent was shown to learn cooperative strategies for defeating enemy agents in single and melee combat. The second approach adopted assumed homogeneous capabilities and responsibilities for the agents, but adjusted their local rewards according to personality preferences. These preferences are modeled after the human psychology instrument called the Myers--Briggs Type Indicator (MBTI). This was implemented in a simulated cooperative resource gathering scenario, and personality type assignment was shown to be an effective way of improving team behavior. Both the reward blending and personality typing the agents are shown to be very effective.;Finally, a method of automatically typing actions according to an information based model was implemented and tested in a MAS comprising 5 Robocode robots, each with a personality-typed Commander, Gunner, and Driver agent. The effect of personality typing on robot teaming in this heterogeneous team was evaluated, and strong team performance sensitivity to MBTI type was discovered for Commander agents. The work has considerable impact on adaptive robot team formation for various applications. Military and biologically inspired cognitive architecture applications are of interest.