Minecraft-DQN-RL

This repository provides and in-depth analysis of several Deep Q-Network (DQN) approaches in terms of performance and hardware footprint. Our target mission is a partially observable navigation problem, modeled in Minecraft, a state-of-the-art training and testing environment for research focusing on lifelong learning. The aim of this work is to compare several approaches fairly on a common task.

Prerequisites

• Python 3.5
• TensorFlow

Installation

• Configure and install the Malmo Platform
• Install Gym-Minecraft

Models included in the analysis

• Double Deep Q-Network
• Stacked Double Deep Q-Network
• Recurrent Double Deep Q-Network
• Dueling Double Deep Q-Network
• Stacked Dueling Double Deep Q-Network
• Recurrent Dueling Double Deep Q-Network

Credits

We would like to thank Clément Romac and Pierre Leroy for their exploratory work on partially observable missions. Tambet Matiisen for his implementation of a flexible and reliable training and testing environment for Minecraft.

Selected references

• Double DQNs:
  • Human-level control through deep reinforcement learning
  • Playing atari with deep reinforcement learning
  • Deep reinforcement learning with double q-learning
• Recurrent DQNs:
  • Deep recurrent q-learning for partially observable mdps
• Dueling DQNs:
  • Dueling network architectures for deep reinforcement learning
• Practical Approach to RL:
  • An introduction to deep reinforcement learning