Policy space identification in configurable environments

Abstract We study the problem of identifying policy space available to an agent in a learning process, having access set demonstrations generated by playing optimal considered space. introduce approach based on frequentist statistical testing identify parameters that can control, within larger parametric After presenting two identification rules (combinatorial and simplified), applicable under different assumptions space, we provide probabilistic analysis simplified one case linear policies belonging exponential family. To improve performance our rules, make use recently introduced framework Configurable Markov Decision Processes, exploiting opportunity configuring environment induce reveal which it control. Finally, empirical evaluation, both discrete continuous domains, prove effectiveness rules.