Selecting concise sets of samples for a reinforcement learning agent

We derive an algorithm for selecting from the set of samples gathered by a reinforcement learning agent interacting with a deterministic environment, a concise set from which the agent can extract a good policy.
The reinforcement learning agent is assumed to extract policies from sets of samples by solving a sequence of standard supervised learning regression problems. To identify concise sets, we adopt a criterion based on an error function defined from the sequence of models produced by the supervised learning algorithm.
We evaluate our approach on two-dimensional maze problems and show its good performances when problems are continuous.