Vision-Based Deep Q-Learning on Simple Control Problems: Stabilization via Neurogenesis Regularization

This paper is concerned with the problem of stabilizing the training of Deep Q-Networks applied to simple Reinforcement Learning problems based on visual inputs. In particular, the paper investigates a recently proposed bioinspired regularization technique, namely adult neurogenesis, where the weights of a random subset of nodes are periodically reset. We compare experimental conditions involving different types of inputs, neural network architectures, and regularization techniques. The experiments reveal that the proposed implementation of adult neurogenesis is capable of effectively speeding up and stabilizing the training process of Deep Q-Networks.