Department of Electrical Engineering
Control Robotics and Machine Learning Lab
Technion - Israel Institute of Technology
המעבדה לבקרה רובוטיקה ולמידה חישובית
Automatic Skill Refinement in Minecraft
using Deep Reinforcement Learning
Background
One of the big challenges in Artificial Intelligence is solving the problem of lifelong learning. An agent operates in a world and is constantly faced with new tasks that it has to solve (similar to humans). The challenge is to learn skills to solve these tasks and then reuse them when a new task, requiring similar skills is encountered. In this case the skills may need to be refined to solve the given task.
Consider the task of building a wooden house. In this case we may learn skills to (1) fetch wood, (2) Lay foundations and (3) place the wood on the foundations to build the house. Now we are faced with a new task, build a brick house. Even though similar skills can be used, we need to refine the fetch wood skill to become a 'fetch bricks' skill etc.
Project Goal
Build on a previous work 'A Deep Hierarchical Approach to Lifelong Learning in Minecraft (published from our group in AAAI 2017). Incorporate automated skill learning and skill refinement into the architecture and test it in Minecraft
Project steps
-
Study and understand our previous Minecraft work
-
Learn about Malmo, the new Minecraft framework developed by Microsoft
-
Get our previous framework to work with Malmo
-
Incorporate skill learning and skill refinement into this framework
Required knowledge
-
Random Signals.
-
Any knowledge in machine learning (DL, RL, t-SNE) is an advantage.
Environment
Comments and links