Leveraging Representation Learning and Expert Execution Traces for Improved Reinforcement Learning Performance

Abstract

Reinforcement Learning (RL) has shown great promise in solving complex decision-making problems. However, the high sample complexity of RL algorithms has motivated researchers to explore alternative approaches that can reduce the amount of data needed for learning. One such approach is to incorporate expert knowledge through supervised learning from human experts. The hypothesis is based on implementing triplet embeddings for supervised learning with human-traced data in Cart Pole, a Reinforcement Learning environment. This thesis explores using representation learning with Siamese networks to extract information from limited expert traces effectively. Experiments in the Cart Pole environment show the effectiveness of this approach. Both the accuracy of the classifiers at choosing the optimal action (label) is measured, as well as their ability to accumulate reward in the domain. The study conclusively established intriguing insights about using triplet embeddings in the Classifier and RL domain.