Deep Reinforcement Learning Agent (DQN algorithm)

This is an implementation of Deep Reinforcement Learning for a navigation task. Specifically, DQN algorithm with experience replay method is used to solve the task.

Details of the environment

THe environment is a Unity Environment which consists of a square surface with Yellow and Blue Bananas scattered around.

Task of the agent

The agent needs to collect as many yellow bananas as possible while avoiding the blue bananas.

Actions possible for agent

• move forward
• move backward
• turn left
• turn right

Reward for the agent

• Yellow banana +1 reward
• Blue banana -1 reward

When is environment said to be solved ?

The banana collection is an episodic game. Idea is to maximise the total score in an episode. The environment is said to be solved if the agent learns to secure an average score of at least +13 points over 100 consecutive episodes.

How to get started ?

• Gain a basic understanding of Unity Environment
• Set up a Python 3.6 Environment to install Dependencies involving PyTorch, the ML-Agent toolkit and a few more Python packages.
• Download a Unity Environment for Windows(64-bit)/Windows(32-bit)/Mac OSX/LINUX

My Solution

• Run Navigation.ipynb
• See a glimpse of my agent during training and my trained agent collecting bananas on YouTube.
• Do checkout my Report for more theoretical explanation of the project implementation.