Boxing Q-Learning

Result when 0% of learning is done.
Result when 50% of learning is done.
Result when 100% of learning is done.

Project information

Project Description
An Agent learns how to box using Naïve Q-Learning and Deep Q Network (DQN).

Introduction

Boxing with Q-Learning was a project for the “Artificial Intelligence” course in which an agent learns how to box using Naïve Q-Learning and Deep Q Network (DQN).

Technical Overview

The project is implemented by Python language with the TensorFlow framework. It is based on Reinforcement Learning and has two phases. In the first phase, implementation was done by the Naïve Q-Learning approach. Since there are many scenarios where tables are not able to scale nicely, in the second phase Deep Q Network is used. It is a deep neural network consisted of five layers: 3 Convolutional and 2 Fully-Connected Layers. With DQNs, instead of a Q Table to lookup values, there is a model that we inference (make predictions from), and rather than updating the Q table, we fit (train) your model. So, the performance and the result will be significantly enhanced. In DQN, the total number of training steps is 5000000.

Architecture of the Network
Network architecture used for the DQN phase

Technologies/Languages Used

Technology Usage
Pytyhon is used for implementing the application
Tensorflow is the framework for implementing the Q-Learning and DQN