Distillation of Deep Reinforcement Learning Models Using Fuzzy Inference Systems

Recently, significant progress has been made in the field of Deep Reinforcement Learning, with advances in a wide variety of application domains, such as arcade game playing, continuous control, and the game of Go. Many of these advances have been made using deep neural networks, which are widely regarded as black boxes. This means that the inner workings of a deep neural network are hard to understand, making interpretation of the learned policy by humans a difficult task. However, interpretability is a critical property of machine learning models in many application domains, including legal and medical applications. In this work, we use policy distillation [1] to distill the learned policy from a deep Q-network to an ANFIS controller [2]. The advantage of neuro-fuzzy controllers such as ANFIS is that they can be trained much like a neural network, but can be much more interpretable. This interpretability is not intrinsic to the model however, and specific precautions must be taken to ensure that the result is as interpretable as possible. For this reason, we extend the original policy distillation algorithm with a pre-processing and a post-processing step to maximize model interpretability [3].