Transparency of Task Dependencies of Reinforcement Learning in Unmanned Systems

Reinforcement learning has been applied widely, however there are yet no easy ways to explain both an agent's behaviour or the complex unmanned systems that an agent is in. In this paper, we design a framework and its corresponding algorithm to discover the hidden rules or relationships in the unmanned systems supported by policies obtained through Reinforcement Learning training. It is demonstrated experimentally that our approach is effective in building an explicit representation for revealing the hidden rules or constraints when an agent needs to accomplish some tasks in that unmanned systems.