Making reinforcement learning more data-efficient

• Main Author: Cao, Jeffery Siming
• Other Authors: Tan Yap Peng
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/149413
• Institution: Nanyang Technological University

Recently, researchers showed that applying deep neural networks to Reinforcement Learning (RL) can improve the agent’s ability to learn action policies directly from high dimensional input states. However, there are still several challenges present when applying these algorithms to real world problems. One huge challenge is that they require a lot of interactions with the environment to learn a suitable action policy and have poor data-efficiency. In supervised learning tasks, data augmentation has been shown to improve the data-efficiency of deep learning models. This paper investigates the effects of data augmentation and other semi-supervised methods on a state-of-the-art RL algorithm, namely the Contrastive Unsupervised Representation for RL (CURL) algorithm. We demonstrate that data augmentation can improve the data-efficiency of the CURL algorithm when applied to the DeepMind Control Suite benchmark running on OpenAI Gym. Furthermore, we investigate possible applications of the Mean Teacher semi-supervised learning method to CURL