基于启发式探测的多智能体分层强化学习

In order to deal with the low performance of random exploration in the early epoches of multi-agent hierarchical reinforcement learning, a multi-agent hierarchical reinforcement learning approach based on heristic exploration is discussed. A heuristic exloration is proposed by combining heuristic Boltzmann machine with roulette algorithm for act selection. In a result, the learning agent converges quickly in early epoches. The simulation expremental results show that the proposed approach is adaptable and advantaged. More expremental results show that heuristic exploration still play an important role.