Related Experiment Video
Updated: Jun 4, 2025
10:56
Long-term Behavioral Tracking of Freely Swimming Weakly Electric Fish
Published on: March 6, 2014
12.5K
Simulating fish autonomous swimming behaviours using deep reinforcement learning based on Kolmogorov-Arnold Networks
Tao Li1,2, Chunze Zhang1, Guibin Zhang1,2
1Southwest Research Institute for Hydraulic and Water Transport Engineering, Chongqing Jiaotong University, Chongqing, People's Republic of China.
Bioinspiration & Biomimetics
|January 3, 2025
Summary
Kolmogorov-Arnold Networks (KANs) improve deep reinforcement learning for simulating fish swimming. KANs offer enhanced perception and decision-making, leading to better performance and faster learning in complex fluid environments.
Area of Science:
- Robotics and Artificial Intelligence
- Biomimetics and Biomechanics
- Computational Fluid Dynamics
Background:
- Fish swimming behavior and locomotion are crucial for scientific and engineering applications.
- Deep Reinforcement Learning (DRL) combined with Computational Fluid Dynamics (CFD) simulates adaptive fish swimming.
- Current DRL models face efficiency challenges in high-dimensional, dynamic environments.
Purpose of the Study:
- To investigate the efficacy of Kolmogorov-Arnold Networks (KANs) in enhancing DRL-based fish swimming simulations.
- To address the need for more efficient network architectures in approximating state-value functions.
- To improve the perception and decision-making capabilities of intelligent agents in complex hydrodynamic environments.
Main Methods:
- Integration of KANs into an existing computational platform for simulating autonomous fish swimming.
- Testing KAN performance in point-to-point and Kármán gait swimming scenarios.
- Comparative analysis against Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) networks.
Main Results:
- KANs significantly improved perception and decision-making in complex fluid environments compared to LSTMs and MLPs.
- KANs achieved substantial average reward improvements: up to 88.0% over MLPs and 94.1% over LSTMs in point-to-point swimming.
- In Kármán gait swimming, KANs showed even greater improvements: 766.7% over MLPs and 105.6% over LSTMs.
- KANs demonstrated faster learning capabilities and more stable swimming performance under comparable network sizes.
Conclusions:
- KANs represent a promising, more efficient network architecture for DRL applications in fluid dynamics and robotics.
- The findings suggest KANs can overcome limitations of traditional networks in complex, uncertain environments.
- This research paves the way for more generalizable and economical simulations of autonomous swimming behaviors.
