Pham_Huy_N.pdf (701.35 kB)
Download file

Applying DTGolog to Large-scale Domains

Download (701.35 kB)
thesis
posted on 22.05.2021, 11:34 by Huy N Pham
While decision theoretic planning (DTP) offers great potential benefits to elicit purposeful behavior of the agent operating in uncertain environments, state-based approaches to DTP are known to be computationally intractable in large-scale domains. DTGolog is a decision-theoretic extension of a logic-based high level programming language Golog that completes a given partial Golog program using a form of directed value iteration. DTGolog has been proposed to alleviate some of the computational difficulties associated with DTP. The main advantages of DTGolog are that a DTP problem can be formulated using a logical representation to avoid explicit state enumeration, and the programmer can encode domain-specific knowledge in terms of high-level procedural templates to partially specify behavior of an agent. These templates constrain the search space to manageable size. Despite these clear advantages, there are few studies that investigate the applicability of DTGolog to very large-scale practical domains. In this thesis, we conduct two studies. First, we apply DTGolog to the well known case-study of the London Ambulance Service to demonstrate advantages and potentials of DTGolog as a quantitative evaluation tool for designing decision making agents. Second, we develop a software interface that allows to control the well-known Sony's AIBO robotics platform using DTGolog. We show that DTGolog can be used on this platform with a minimal amount of software customization. We run experiments to test functionality of our interface. The main contribution of this thesis is demonstration of applicability of DTGolog to two different large scale domains that are both practical and interesting.

History

Language

eng

Degree

Master of Applied Science

Program

Electrical and Computer Engineering

Granting Institution

Ryerson University

LAC Thesis Type

Thesis

Thesis Advisor

Mikhail Soutchanski