Training of an agent playing Super Mario through RL and CNN

Introduction

This project is an implementation of a Deep Q-Network (DQN) agent for reinforcement learning tasks. The agent is designed to interact with an environment and learn an optimal policy over time.

Features

• Customizable hyperparameters: The agent's learning rate, discount factor, epsilon (for epsilon-greedy action selection), epsilon decay rate, minimum epsilon, batch size, and the rate at which the target network is updated can all be customized.
• Replay buffer: The agent uses a replay buffer to store past experiences, which are then sampled in batches to update the network.
• Epsilon-greedy action selection: The agent uses an epsilon-greedy policy for action selection, with epsilon decaying over time to a specified minimum value.

Usage

The main file is the MARIO.ipynb file.

To use the agent, you can create an instance of the agent with your desired hyperparameters, then call the agent's methods to interact with your environment and update the agent's networks.

Hyperparameters

• input_dims: The dimensions of the input to the neural network.
• num_actions: The number of actions the agent can take.
• lr: The learning rate for the network updates.
• gamma: The discount factor for future rewards.
• epsilon: The initial probability of the agent taking a random action.
• eps_decay: The rate at which epsilon decays over time.
• eps_min: The minimum value to which epsilon can decay.
• replay_buffer_capacity: The capacity of the replay buffer.
• batch_size: The size of the batches sampled from the replay buffer to update the networks.
• sync_network_rate: The number of steps that should pass before the target network is updated to match the main network.