Dynamic Partition of Collaborative Multiagent Based on Coordination Trees

In team Markov games research, it is difficult for an individual agent to calculate the reward of collaborative agents dynamically. We present a coordination tree structure whose nodes are agent subsets or an agent. Two kinds of weights of a tree are defined which describe the cost of an agent collaborating with an agent subset. We can calculate a collaborative agent subset and its minimal cost for collaboration using these coordination trees. Some experiments of a Markov game have been done by using this novel algorithm. The results of the experiments prove that this method outperforms related multi-agent reinforcement-learning methods based on alterable collaborative teams.