Autonomous Discovery of Abstractions through Interaction with an Environment

The ability to create and use abstractions in complex environments, that is, to systematically ignore irrelevant details, is a key reason that humans are effective problem solvers. My research focuses on using machine learning techniques to enable greater autonomy in agents. I am particularly interested in autonomous methods for identifying and creating multiple types of abstractions from an agent’s accumulated experience in interacting with its environment. Specific areas of interest include: Knowledge representation. How can we efficiently represent the knowledge learned in one task and reuse it for other tasks? This knowledge can take the form of a control policy learned to solve one task or a representation of structure in an environment. Autonomous discovery of structure. My dissertation [1] focuses on autonomously identifying and creating useful temporal abstractions from an agent’s interaction with its environment. Interaction of reinforcement learning and supervised learning methods. I am particularly interested in the combined use of these techniques to create more robust and autonomous learning systems. Application of these techniques to robots, with a particular focus on robots assisting a human presence in space.