Nathaniel Hamilton

Abstract: Reinforcement Learning (RL) has become an increasingly popular subject as the success of these algorithms and methods have grown. Concerns regarding the safety of unconstrained agents has led to increased interest in work concerning Safe Reinforcement Learning. While these Safe RL methods provide safer, more reliable and trustworthy policies, they often suffer from a decrease in performance and longer training times. In this work, we explore how a data augmentation framework, Kaleidoscope Experience Replay, can offset these setbacks. We implement these modifications on two well-known deep RL algorithms, Deep Deterministic Policy Gradient (DDPG) and Soft Actor-Critic (SAC), and compare their impact when training in the Pendulum-v0 environment from OpenAI Gym.

Biography:

Bio: Nathaniel Hamilton is a National Defense Science and Engineering Graduate (NDSEG) Fellow currently completing his PhD studies at Vanderbilt University in Nashville, TN and plans to graduate in Spring 2022. His research focuses in Safe Reinforcement Learning and how we can train safer, more robust learning-enabled cyber-physical systems. The past three summers, he has interned with the Air Force Research Lab’s Autonomy Technology Research Center and worked with members of the Autonomy Capability Team (ACT3) on various projects.