A Minecraft Agent Based on Hierarchical Deep Reinforcement Learning Model

A deep reinforcement learning model named MineCLIP has been proposed for solving a given task in imperfect information game Minecraft by learning to associate gameplay videos corresponding to linguistic goals. The goal of the present research is to improve the success rate of the task by applying a hierarchical reinforcement learning framework to the linguistic guidance in MineCLIP, which decomposes the linguistic goal of a given task into steps. MineCLIP uses policy gradient PPO as a reinforcement learning algorithm. The proposed method uses a hierarchy of PPOs in MineCLIP to assign a single sub-policy to each language guidance, which allows step-by-step learning and facilitates task success. Using Minecraft simulation environment, i.e. MineDojo, we confirmed that the proposed method can improve the success rate of solving tasks compared to MineCLIP.