Inverse reinforcement learning with evaluation

摘要

Reinforcement learning (RL) is a method that helps programming an autonomous agent through human-like objectives as reinforcements, where the agent is responsible for discovering the best actions to fulfil the objectives. Nevertheless, it is not easy to disentangle human objectives in reinforcement like objectives. Inverse reinforcement learning (IRL) determines the reinforcements that a given agent behaviour is fulfilling from the observation of the desired behaviour. In this paper we present a variant of IRL, which is called IRL with evaluation (IRLE) where instead of observing the desired agent behaviour, the relative evaluation between different behaviours is known by the access to an evaluator. We present also a solution for this problem under the assumption that a relative linear function that preserves the order assumed by the evaluator exists and that the evaluator evaluates policies instead of behaviours. This is posed as a linear feasibility problem, whose solution is well known. Results of simulations of a set of heterogeneous robots in a search and rescue scenario are presented to illustrate the method and the possibility to transfer the learned reinforcement function among robots