Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Author ORCID Identifier

N/A

AccessType

Open Access Dissertation

Document Type

dissertation

Degree Name

Doctor of Philosophy (PhD)

Degree Program

Computer Science

Year Degree Awarded

2015

First Advisor

Andrew G. Barto

Subject Categories

Artificial Intelligence and Robotics | Robotics

Abstract

One of the defining characteristics of human intelligence is the ability to acquire and refine skills. Skills are behaviors for solving problems that an agent encounters often—sometimes in different contexts and situations—throughout its lifetime. Identifying important problems that recur and retaining their solutions as skills allows agents to more rapidly solve novel problems by adjusting and combining their existing skills. In this thesis we introduce a general framework for learning reusable parameterized skills. Reusable skills are parameterized procedures that—given a description of a problem to be solved—produce appropriate behaviors or policies. They can be sequentially and hierarchically combined with other skills to produce progressively more abstract and temporally extended behaviors. We identify three major challenges involved in the construction of such skills. First, an agent should be capable of solving a small number of problems and generalizing these experiences to construct a single reusable skill. The skill should be capable of producing appropriate behaviors even when applied to yet unseen variations of a problem. We introduce a method for estimating properties of the lower-dimensional manifold on which problem solutions lie. This allows for the construction of unified models for predicting policies from task parameters. Secondly, the agent should be able to identify when a skill can be hierarchically decomposed into specialized sub-skills. We observe that the policy manifold may be composed of disjoint, piecewise-smooth charts, each one encoding solutions for a subclass of problems. Identifying and modeling sub-skills allows for the aggregation of related behaviors into single, more abstract skills. Finally, the agent should be able to actively select on which problems to practice in order to more rapidly become competent in a skill. Thoughtful and deliberate practice is one of the defining characteristics of human expert performance. By carefully choosing on which problems to practice the agent might more rapidly construct a skill that performs well over a wide range of problems. We address these challenges via a general framework for skill acquisition. We evaluate it on simulated decision-problems and on a physical humanoid robot, and demonstrate that it allows for the efficient and active construction of reusable skills.

DOI

https://doi.org/10.7275/6439644.0

Share

COinS