CartPole Deep Q-Network
Explored deep reinforcement learning on the CartPole-v1 environment, investigating how experience replay influences training performance and stability.
Overview
An experimental deep reinforcement learning project exploring Deep Q-Networks (DQN) in the CartPole-v1 environment.
The objective is to train an agent to balance a pole by learning which actions maximize its cumulative reward.
Methodology
The project investigates the effects of experience replay on DQN training.
Experiments included:
- Training a DQN agent on CartPole-v1.
- Incorporating experience replay into the learning process.
- Exploring variations in replay frequency.
- Comparing training rewards across different experimental configurations.
Observations
The experiments suggested that experience replay improved training rewards compared with the baseline approach.
Some variations occasionally performed better, although the results still exhibited instability.
Potential improvements identified included adjusting the target network and experience replay configuration to produce more consistent training.
Technologies
Implemented in Python using a Jupyter notebook.