Meta-Learning for Planning: Automatic Synthesis of Sample Based Planners

Lucas Paul Saldyt, Heni Amor

Keywords: learning, planning, meta-learning, genetic programming, python, machine learning, artificial intelligence

TL;DR: This paper describes methods for learning a diverse array of specialized path planning algorithms using pareto evolution (multi-objective genetic program learning)

Abstract: In this paper, we discuss the challenge of generating domain-specific path planners in a data-driven fashion. Via the multi-objective optimization of Python code, we synthesize new sampling-based path planners that allow robots to adapt to new tasks and environments involving sequential decision-making. In addition to the ability to adapt to new environments, our approach also enables robots to balance their computational needs with improvements in task performance. We show that new computer programs can be generated which represent diverse variants of RRT* optimized to StarCraft maps.

Proposed Reviewers:

Hao Zhang, [e-mail]
Yunyi Jia, [e-mail]