This paper introduces a novel structure based on Kolmogorov-Arnold Networks (KANs) in reinforcement learning tasks. The adaptive and learnable activation function replaces the B-spline function, enhancing the model’s expressivity and yielding a compact structure during training. To address challenges associated with KAN’s inherent recursiveness, we integrate exponentially decaying sinusoid functions, which also enhance GPU optimization and offer an interpretable alternative to standard RL networks. Our implementation achieves comparable performance while using fewer parameters than a standard neural network, thereby reducing memory consumption. Additionally, a parameter reduction in the global stream contributes to smoother learning dynamics and a more streamlined network design. This innovation not only minimizes the learning size but also enables efficient deployment on resource-constrained devices.