Complexification through gradual involvement and reward Providing in deep reinforcement learning

Abstract

Training a relatively big neural network within the framework of deep reinforcement learning that has enough capacity for complex tasks is challenging. In real life the process of task solving requires system of knowledge, where more complex skills are built upon previously learned ones. The same way biological evolution builds new forms of life based on a previously achieved level of complexity. Inspired by that, this work proposes ways of increasing complexity, especially a way of training neural networks with smaller receptive fields and using their weights as prior knowledge for more complex successors through gradual involvement of some parts, and a way where a smaller network works as a source of reward for a more complicated one. That allows better performance in a particular case of deep Q-learning in comparison with a situation when the model tries to use a complex receptive field from scratch.