Machine Learning for Automated Planning

Abstract

Automated planning is a long-standing problem which concerns finding an action sequence to solve a task. In this thesis, we explore two problems in leveraging machine learning for automated planning: (1) learning from failed planning attempts to improve efficiency of future planning, and (2) adding goal-conditioning to action samplers in a neuro-symbolic planning framework. In both problems, we utilize neural networks to learn mappings that empirically enhance the performance of existing planning frameworks.