Symmetry Detection and Exploitation for Function Approximation in Deep RL.

Anuj Mahajan, Theja Tulabandhula

Abstract: With recent advances in the use of deep networks for complex reinforcement learning (RL) tasks which require large amounts of training data, ensuring sample efficiency has become an important problem. In this work we introduce a novel method to detect environment symmetries using reward trails observed during episodic experience. Next we provide a framework to incorporate the discovered symmetries for functional approximation to improve sample efficiency. Finally, we show that the use of potential based reward shaping is especially effective for our symmetry exploitation mechanism. Experiments on classical problems show that our method improves the learning performance significantly by utilizing symmetry information.