We identify a specialised class of reinforcement learning problem in which the agent(s) have the goal of gathering information (identifying the hidden state). The gathered information can affect rewards but not optimal behaviour. Exploiting this characteristic, an algorithm is developed for evaluating an agent's policy against all possible hidden state histories at the same time. Experimental results show the method is effective in a two-dimensional multi-pursuer evader searching task. A comparison is made between identical policies, joint policies and "relational" policies that exploit relativistic information about the pursuers' positions.
展开▼