Adaptive Agents in Minecraft: A Hybrid Paradigm for Combining Domain Knowledge with Reinforcement Learning

摘要

We present a pilot study focused on creating flexible Hierarchical Task Networks that can leverage Reinforcement Learning to repair and adapt incomplete plans in the simulated rich domain of Minecraft. This paper presents an early evaluation of our algorithm using simulation for adaptive agents planning in a dynamic world. Our algorithm uses an hierarchical planner and can theoretically be used for any type of 'bot'. The main aim of our study is to create flexible knowledge-based planners for robots, which can leverage exploration and guide learning more efficiently by imparting structure using domain knowledge. Results from simulations indicate that a combined approach using both HTN and RL is more flexible than HTN alone and more efficient than RL alone.