Improving Deep Reinforcement Learning via Transfer

While achieving the state-of-the-art performance in complex sequential tasks, deep reinforcement learning (deep RL) remains extremely data inefficient. Many approaches have been studied to improve the data efficiency of deep RL algorithms. This dissertation focuses on leveraging various transfer learning techniques to tackle this problem. We first show that positive transfer can be achieved cross-domains via direct weight transfer if the two agents share a certain amount of similarities. Then we look into how could the similarity between cross-domain tasks be quantified, such that we only transfer useful information from one task to another while blocking information that might have a negative effect. The third direction we studied is the human-agent transfer mechanism, which we integrate human knowledge via supervised pre-training on a set of demonstration data collected from a human then transfer to an agent. Lastly, several future directions are proposed for the remainder of this dissertation.