This project has been implemented with the gymnasium Framework: https://gymnasium.farama.org/environments/atari/freeway/
The final selected model was v_20.2
-
Clone the Repository:
git clone https://github.com/JanMuehlnikel/Atari-Freeway-Reinforcement-Learning cd your-repo
-
Create and Activate a New Conda Environment:
conda create --name FreewayEnv python=3.10.14 conda activate FreewayEnv
-
Install
pip
in the New Conda Environment: @@ -27,39 +27,13 @@ This project has been implemented with the gymnasium Framework: https://gymnasiupip install -r requirements.txt
This project implements a reinforcement learning agent using a Deep Q-Network (DQN). Despite our efforts, the agent was only able to learn the baseline method, which involved simply moving up.
- The baseline method for the agent was to only move up.
- Algorithm: Deep Q-Network (DQN)
- Baseline Method: Move up
- Reward Structure: The reward might have been too high for moving actions and too low for crashing, affecting the learning process.
- Network Complexity: The neural network used might have been too simple to accurately approximate the Q-values, limiting the agent's ability to learn more complex behaviors.
- Training Environment: Training was conducted on a CPU with a limited training period, which may have constrained the agent's learning capability.
Several factors might have contributed to the agent's limited learning:
- Adjusting Rewards: Modifying the reward structure to balance rewards for moving and penalties for crashing.
- Network Architecture: Using a more complex neural network to better approximate Q-values.
- Training Resources: Utilizing a GPU for training and extending the training period to improve learning efficiency.
These adjustments could potentially enhance the agent's ability to learn beyond the baseline method.