Training AI to play Atari​ using OpenAI Gym and Deep Reinforcement Learning

For this task, I used the openai gym framework that renders the atari environment and exposes APIs​ to the game’s world.

My agent was a Deep Q-learning agent that takes in sequences of 4 images of the game’s world as inputs and learns the Q functions for several actions in any given state (sequence of 4 images of the game). These actions include left, right, fire.

The agent used a deep learning architecture consisting of 3 convolution layers and one dense layer. The agent trained using target networks and experience replay buffers to prevent the propagation of errors in learning across its deep learning stack.