A New Sample-Efficient PAC Reinforcement Learning Algorithm

This paper introduces a new hybrid PAC RL algorithm for MDPS, which intelligently maintains favorable features of its parents. The DDQ algorithm, integrates model-free and model-based learning approaches, preserving some advantages from both. A PAC analysis of the DDQ algorithm is presented and its sample complexity is explicitly bounded. Numerical results from a small-scale example motivated by work on human-robot interaction models corroborates the theoretical predictions on sample complexity.