Solving Long-Horizon Multi-Step Tasks by Combining Planning and Deep Reinforcement Learning

Abstract:

While planning methods are robust for long horizon prob-lems in domains that can be easily and effectively expressedin a symbolic representation, these methods do not adapt wellto uncertainties in dynamic environments. Similarly, for tra-ditional Deep Reinforcement Learning (DRL) problems, thedata availability becomes a crucial bottleneck in solving com-plex problems. Furthermore, the black-box nature of DRL so-lutions affects the interpretability of the approach. The ideabehind hybrid approaches that combine both planning and re-inforcement learning is to leverage the benefits of both themethods for efficient and robust decision-making in complex,and often, evolving environments. In this project, we explorethis concept, formulate aplanning+RLframework for twowidely used domains - Taxi and the Atari game: Montezuma’sRevenge, and report insights from our experiments.

Note: Experimental results will be added soon.

References:

1. [Lyu et al. 2019] Lyu, D.; Yang, F.; Liu, B.; and Gustafson,S. 2019. Sdrl: interpretable and data-efficient deep rein-forcement learning leveraging symbolic planning. InPro-ceedings of the AAAI Conference on Artificial Intelligence,volume 33, 2970–2977. Github Link: https://github.com/daomingAU/MontezumaRevenge_SDRL