EECS – Defense: Bootcamp Method for Training General Purpose AI Agents, Vincent Lombardi

About the event

Student: Vincent Lombardi

Advisor: Dr. Lawrence Holder

Degree: Computer Science MS

Thesis Title: Bootcamp Method for Training General Purpose AI Agents

Abstract: General purpose agents have long been an ultimate goal of AI research. One promising approach to this goal is to first train an agent to use a variety of skills, called a skillnet agent, and then allow the agent to learn how to choose the appropriate skill instead of having to choose the appropriate low-level action. However, there are multiple ways in which skills can be trained which impacts how well the agent retains earlier skills. We propose a method for training skillnet agents called Bootcamp that helps agents efficiently learn skills that can be used to solve a variety of tasks. Bootcamp is a method of training a skill-based hierarchical reinforcement learning agent to learn basic skills in an environment. We found that Bootcamp agents outperform skillnet networks trained randomly on various tasks defined in the ViZDoom simulated environment. We also found that skillnet agents outperform more conventional reinforcement-based learning approaches such as DQNs in ViZDoom.