Policy Space Identification in Configurable Environments

TL;DR

This paper introduces statistical testing methods to identify the controllable policy parameters of an agent within configurable environments, enhancing understanding of agent capabilities through probabilistic analysis and empirical validation.

Contribution

It presents novel identification rules for policy space detection, including a probabilistic analysis for linear policies, and leverages environment configurability to improve identification accuracy.

Findings

Effective policy space identification in discrete and continuous domains

Probabilistic analysis validates the simplified identification rule

Configurable environments enhance parameter control detection

Abstract

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