Content area
Heuristic search is a universal problem-solving method and the core of automated planning, which offers key theoretical properties such as completeness and solution quality guarantees, including w-optimality.
Although machine learning (ML) has experienced significant success in areas like image classification and natural language processing, its application to search and automated planning presents considerable challenges. Current methods frequently sacrifice the theoretical benefits of traditional search algorithms, underscoring the need for approaches that merge the adaptability of ML with the dependability of search-based techniques.
This thesis explores how ML methods can be effectively integrated into heuristic search to enhance efficiency and scalability without sacrificing critical guarantees. It addresses three research questions: (1) How can learned heuristics and policies be incorporated into search algorithms while maintaining suboptimality guarantees? (2) How can GPU-based batch processing accelerate the computation of learned heuristics without compromising these guarantees? (3) How can partially specified symbolic models enhance the performance of black-box search algorithms using ML-based heuristics or policies?
Key contributions include the development of Focal Discrepancy Search (FDS), a bounded suboptimal search algorithm that integrates learned heuristics while preserving theoretical guarantees, and K-Focal Search, a GPU-accelerated method that improves search efficiency while providing w-optimal solutions. Furthermore, this work introduces a novel framework that combines partial symbolic models with ML-based approaches, enabling more scalable and efficient solutions for sequential decision-making tasks.