Learning spatial and expertise distribution coordination in multiagent systems

The reinforcement learning (RL) strategy used in a single autonomous agent needs to be reconsidered and adapted to the multiagent environment. This implies that the strategy adopted should take into account the behavior of other agents to learn to perform effectively in a cooperative manner. In this paper, we present a coordination method based on the RL, which enables an agent to learn from observing the behavior of other agents and their effect on the common environment. The agents coordinate their actions by including other agents' location information within their visual field into their state vector. In this case, the agent is "aware" of its team as a form of geographical distribution. In addition, the reinforcement signal is redefined to include portions of the rewards gained by nearby agents. Thus, each agent acts in a social way and the groups of agents with this reshaping reinforcement signal are learning to behave cooperatively by avoiding conflicts through expertise distribution.