KAN-Agents is a deep reinforcement learning repository implementing Kolmogorov-Arnold Networks (KAN) with various RL agents. The project integrates KAN with reinforcement learning algorithms like PPO and SAC and benchmarks them across different environments.
- Proximal Policy Optimization (PPO)
- KAN-PPO (PPO with Kolmogorov-Arnold Network)
- Soft Actor-Critic (SAC) - [Work in Progress]
- Decision Transformer (DT) - [Work in Progress]
The following environments have been successfully trained using PPO and KAN-PPO:
CartPole-v1LunarLander-v3HumanoidStandup-v5Humanoid-v5InvertedDoublePendulum-v5InvertedPendulum-v5Walker2d-v5
Below is the performance comparison of PPO and KAN-PPO across different environments. The results, including sample efficiency, final performance, and training stability, are stored in the comparison folder.
| CartPole | HumanoidStandup |
|---|---|
 |
 |
| Humanoid | InvertedDoublePendulum |
| --------- | ------------- |
 |
 |
| InvertedPendulum | LunarLander |
| --------- | ------------- |
 |
 |
| InvertedPendulum | |
| --------- | |
 |
KAN-Agents/
│── agents/ # PPO, KAN-PPO implementations
│── comparison/ # Comparison results and analysis
│── results/ # Performance metrics and logs
│── train.py # Training scripts for agents
│── test.py # Evaluation scripts
│── README.md # Project documentation
-
Clone the repository:
git clone https://github.com/MorningStarTM/KAN-Agents.git cd KAN-Agents -
Install dependencies:
pip install -r requirements.txt
- SAC with KAN for continuous control environments
- Integrate DreamerV3 for model-based RL
- Custom-designed environments for real-world applications
This project is released under the MIT License.
⭐ Star this repository to stay updated on future developments!
This project was inspired by the work of Blealtan, who implemented the Efficient Kolmogorov–Arnold Networks (KAN) Layer. You can check out their original repository here: