Department of Electrical Engineering Technion - Israel Institute of Technology
Control Robotics and Machine Learning Lab- המעבדה לבקרה רובוטיקה ולמידה חישובית
Multi level games in Atari
Background
The Arcade Learning Environment (ALE) provides an interface to hundreds of Atari 2600 game environments, each one different, interesting, and designed to be a challenge for human players.
ALE presents significant research challenges for reinforcement learning such as the well known "Curse of dimensionality".
One way to tackle that is by approximating the value function using Deep Networks for example.
Project Goal
Many ALE games have a number of levels that vary in difficulty. Once a player finishes a level, they begin playing on a new level. Two examples are Breakout and Qbert:
-
Breakout - once the player has finished clearing the first screen he receives an identical second screen which he has to clear.
-
Qbert - The game is run in rounds. At each round the colors of the screen change but the goal remains the same.
In this project you will first characterize aspects of multi level games that inhibit the learning capabilities of the DQN agent. You will then suggest algorithms to try and overcome these difficulties.
Project steps
-
Gain a basic understanding of reinforcement learning and deep learning.
-
Run google DQN(Deep Q-Learning) code, and examine multi level games. Characterize the aspects that inhibit the agent's learning.
-
Suggest algorithms to deal with these problems.
Required knowldege
-
Random Signals
-
Any knowledge in machine learning(DL,RL) is an advantage.
Enviroment
Comments and links
