[Proposal-ML] Training an Agent to find Diamond in Minecraft, MineRL Competition 2021
Keywords: Reinforcement Learning, Imitation Learning, Minecraft, MineRL
Abstract: This proposal presents a novel approach for training an agent to solve the "Obtain Diamond" task in Minecraft, as defined by the MineRL competition 2021. The competition poses a unique challenge, as the agent must navigate a randomly generated environment to accomplish a series of complex tasks before successfully mining a diamond. Our method combines Hierarchical Reinforcement Learning (HRL) with Curriculum Learning, incorporating human demonstrations to enhance sample efficiency. Specifically, we adopt a "Top-Down" or "reverse" Curriculum Learning strategy, training the agent to accomplish the goal from progressively simpler sub-goals, starting near the objective and working backward to the starting state. By leveraging both human data and a structured hierarchy of goals, our approach aims to increase efficiency and adaptability, making it well-suited to handle the complex nature of the Minecraft environment.
Submission Number: 54
Loading