Project steps
-
Gain basic understanding of RL and DL.
-
Understand and run google's Deep Q-Learning(DQN) code.
-
Extend DQN with RL Options and test on Seaquest.
Required knowledge
-
Random Signals.
-
Either knowledge in RL or motivation to study are necessary.
-
Any knowledge in machine learning(DL,RL) is an advantage.
Enviroment
Comments and links
-
In order to understand the options frmework only basic understanding of RL is required.
-
The project is relevant in various research aspects and may lead to publication.
Department of Electrical Engineering Technion - Israel Institute of Technology
Control Robotics and Machine Learning Lab- המעבדה לבקרה רובוטיקה ולמידה חישובית
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 using aproximations for the value function such as Deep Networks.
Project Goal
In Seaquest you play a sub-marine in a shooting game. What makes the game more complicated is the oxygen bar. The player's oxygen levels decrease as time increases. In order to refill his oxygen tanks, the player needs to surface at sea level.
Reinforcement Learning(RL) typically contains single time step actions e.g. move left/right. These actions often cannot explore sufficiently in high dimensional domains(like games). Therefore multi-time step actions called options have been introduced into RL to more efficiently explore large domains and provide better solutions.
In this project we aim to learn better RL policies using the options framework.