top of page
האתר הזה עוצב בעזרת הכלי לבניית אתרים של
.com
. רוצים ליצור אתר משלכם?התחילו עכשיו
Department of Electrical Engineering
Control Robotics and Machine Learning Lab
Technion - Israel Institute of Technology
המעבדה לבקרה רובוטיקה ולמידה חישובית

Successor features for Minecraft

Background
 

Minecraft provides a challenging framework for RL, since it is a high dimensional and stochastic domain. It is possible to construct your own environment (e.g. a maze) and train the agent to traverse it. You can do anything in Minecraft from building houses to creating farms.

 

Successor features: Our focus is on transfer where the reward functions vary across tasks while the environment's dynamics remain the same.This allows a value function representation that decouples the dynamics of the environment from the rewards and allows transfer. 

Project Goal

 

In this project, we will examine a transfer learning framework for Minecraft, where only the reward signal changes between different tasks. The goal is to use as much knowledge as possible from previous tasks to learn new tasks. We will build on previous results on Successor representations in RL and examine how the method performs in Minecraft. 

Project steps

 

  • Understand the Minecraft game and framework

  • Understand the DQN framework

  • Get a basic understand of the Successor features literature

  • Learn a set of tasks in the same domain without transfer

  • Investigate different transfer methods and compare with successor representations

 

Required knowledge

 

  • Strong programming skills. 

  • Any knowledge in DL and RL is an advantage. 

 

Environment

  • Torch/TensorFlow.

 

Comments and links

 

  • See the DQN paper by Google

  • See this paper regarding successor features and look for follow-up works.

 

 

bottom of page