CAREER: Solving Pathfinding Problems with High-Level Goal Specifications

Pathfinding is essential for making effective decisions about how to move through complex spaces, from robots navigating crowded warehouses to synthesizing molecules through sequences of reactions. A significant roadblock to solving these pathfinding problems is that the goal may only be describable at a high level. For example, in chemical synthesis, scientists may know the properties a therapeutic molecule should have, but not know of any molecule that has such properties. This project seeks to remove this roadblock by creating novel artificial intelligence (AI) algorithms to solve pathfinding problems based on high-level goal descriptions. This could lead to the discovery of new goal configurations and reduce the time and cost required to solve pathfinding problems. The objective of this project is to address high-level goal specification and goal reaching for pathfinding problems, where a high-level goal specification defines a set of goal states without explicitly specifying any state in the set. The ability to find paths based on high-level goal specifications becomes necessary when properties of a goal state can be specified, but states that satisfy these properties are not known. This project will develop novel domain-independent AI algorithms that combine machine learning, heuristic search, and formal logic to reach goals specified with answer set programming, an expressive logic programming language. Deep reinforcement learning will be used to train a deep neural network heuristic function that estimates the distance between a given state and a high-level goal specification defined using answer set programming. To learn from failures, a neuro-symbolic algorithm that refines the heuristic function using symbolic constraints extracted from failure cases will be developed. The project will also contribute to education and training of students in AI by providing hands-on experience, exposure to real-world applications, and engagement with current research challenges that build both practical skills and deeper conceptual understanding. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2541764 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Forest Agostinelli | Institution: University of South Carolina at Columbia, COLUMBIA, SC | Award Amount: $282,156 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2541764 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2541764.html