Online Resource Optimization and Computation Offloading in Edge Networks with KANH-PPO

Traditional reinforcement learning methodologies, primarily based on multi-layer perceptron (MLP) architectures, require extensive, fully connected layers for complex nonlinear representations. Such an approach increases computational demands and enhances the likelihood of model overfitting. In this paper, we present an innovative reinforcement learning approach, leveraging the Kolmogorov-Arnold Networks (KAN) framework, to enhance decision-making processes and resource management strategies in edge networks. We develop a KAN-based hybrid proximal policy optimization algorithm (KANH-PPO) to address this issue. This algorithm effectively addresses the challenge of hybrid action spaces within edge networks, which include discrete action spaces characterized by computational offloading decisions and continuous action spaces characterized by power allocation. Furthermore, the KANH-PPO algorithm innovatively integrates the KAN architecture, significantly reducing the number of training parameters and enhancing the algorithm's fitting capability and overall performance. Simulation experiments indicate that our proposed KANH-PPO algorithm outperforms the benchmark algorithm in terms of convergence speed and edge network system power, and it requires significantly fewer training parameters than benchmark algorithms. This helps to reduce the power consumption of communication and promote the development of green communication.