Transfer learning for deep neural nets

Abstract

The objective of this master thesis is to find transfer learning methods which can be applied in reinforcement learning problems. This method has to improves the learning speed of an agent on a target task, given a set of already trained policies on similar simpler tasks. We used Q learning algorithm combined to neural network to represent Q function that determine how an agent act on it’s assigned task. Using source network trained on simpler problems, we manage to find an architecture that combines networks with an interface to transfer knowledge in the new network. This technique allowed us to solve the OpenAI Gym mountain car problem using source networks trained on a smaller mountain and a faster car. We tested more complex connection architectures to improve the representation complexity of the combined networks. We also implemented a method to learn multiple networks in parallel to exploit the ensemble properties when learning a new task. This method improves the convergence and results. Finally we tested our implementation on the OpenAI Gym Cartpole environment with succes.