Department of Electrical Engineering
Control Robotics and Machine Learning Lab
Technion - Israel Institute of Technology
המעבדה לבקרה רובוטיקה ולמידה חישובית
Natural Language State Space Representation
Background
Recent advances in Reinforcement Learning have highlighted the difficulties in learning within complex high dimensional domains.
We argue that one of the main reasons that current approaches don't perform well is due to insufficient representation techniques.
A natural way of describing what we observe is through natural language, e.g., "you are inside a room, to your right is a monster".
Project Goal
-
In this project, you will build on top of the game of Doom and create a complex task - navigation, item collection, and more.
-
The novelty of your work is that you will introduce a new way of representing the states - using natural language.
-
You will compare your method to the standard approach of learning from visual inputs.
-
Suggest and implement your ideas!
This project uses state of the art approaches thus hard work may lead to a publication at a top venue.
Project steps
-
Understand the Doom domain.
-
Understand Reinforcement Learning training methodology.
-
Introduce your new and complex domain in Doom.
-
Implement a natural language descriptor module.
-
Train an RL agent on both your new representation and the standard visual one and compare.
-
Suggest your own ideas!
Required knowledge
-
Strong programming skills.
-
Experience with Python is a must.
-
Knowledge in Deep Learning.
-
Any knowledge in RL is an advantage.
Environment
-
Python, PyTorch / TensorFlow.
Comments and links
-
See the DQN paper by Google.