Learning Abstractions for Generalized Reasoning in AI
Generalization is an important ability that allows humans to tackle complex problems by identifying common problem structures and the essential details to do reasoning. Although this ability comes naturally to humans, providing it in AI systems turned out to be a challenge. Constructing “good” abstractions over the problems from a few instances is key to finding generalized solutions that work for all potential instances with similar structure. The obtained generalized solutions can also be used as a guide in searching for a solution to the original problem. Machine learning methods excel at certain generalization tasks, but the generated models lack transparency and guarantees of the generalization. In Symbolic AI, different approaches have been proposed for simplification, abstraction of details, and generalization, usually focusing on concrete domains. However a general theoretical understanding of the generalization ability, and more importantly, methods to compute abstractions, especially those that are domain-independent, are still lacking. This project takes on this challenge in the context of answer set programming (ASP), one of the premier formalisms of symbolic AI. ASP is a modern declarative problem solving approach which, thanks to its efficient solvers and expressive language, is increasingly popular for solving a variety of problems in AI and Computer Science. With a strong team, we aim to establish theoretical foundations for generalized reasoning, and to develop methods for learning useful abstractions over programs that allow one to obtain generalized solutions. The experimental implementations will be evaluated with respect to performance and solution quality on industrial-type problems and generalization tasks. Our results will pave the way for a new generation of solvers with generalization abilities that are applicable to various problems.