Generalization in Planning

The ability to generalize from past experiences in order to address new situations is one of the cornerstones of human intelligence. As such, the design of artificial intelligence systems must---at some point---consider the issues involved in generalization. In automated planning, where the objectives are to design systems that are capable of finding courses of action to achieve specific goals given a formal description of their environment, we can say that a system exhibits generalization capabilities if it can leverage its previous solution-finding efforts when addressing new problems. The pervasive use of automation in modern industries signifies that this type of generalization---generalization in planning---is a fundamental requirement for the integration of artificial intelligence techniques into real-world applications. This dissertation aims to provide a generic high-level approach that can be used when confronted with sequential decision-making problems where generalization is important. Overall, the approach works by reformulating the problems into abstract representations, finding solutions for these abstract problems, and attempting to directly use the resulting abstract solutions in the concrete problems. This satisfies the generalization requirements when multiple concrete problems can be represented in one single abstract problem, or when the insights gathered when solving one abstract problem can be transferred towards solving related problems. We instantiate this approach in three different classes of planning problems. First, we address planning problems that have propositional and numeric state variables, including problems with nonlinear numeric constraints. We then address a type of generalized planning where a family of multiple planning problems is described through the use of first-order logic quantification, obtaining a single policy that can be applied on any of the problems. Finally, we show how we can solve reinforcement learning problems where multiple different tasks must be solved in a single environment. As with all reinforcement learning problems, the exact domain dynamics are initially unknown, and solutions must be obtained by repeatedly interacting with the environment. We prove the soundness of all our approaches and present empirical results that demonstrate their efficacy across various different domains. In many cases, we see orders of magnitude improvements in overall time efficiency.

Department dc:contributor.department

Computer Science

Year dc:date.issued

2022

Author dc:creator

• Illanes, León

Advisor dc:contributor.advisor

• McIlraith, Sheila Ann

• Abstraction
• Automated Planning
• Generalized Planning
• Hierarchical Reinforcement Learning
• Numeric Planning
• Reinforcement Learning

Handle dc:identifier.uri

http://hdl.handle.net/1807/123532

OAI identifier oai:identifier

oai:utoronto.scholaris.ca:1807/123532