Hybrid optimization for efficient pruning of randomly generated neural networks

Randomly generated neural networks, such as reservoir computing (RC) and extreme learning machines (ELMs), simplify training by fixing randomized weights, but suffer from computational redundancy and high hardware demands due to their reliance on excessive neurons. This paper proposes a hybrid regularization framework integrating L1 and L2 optimization strategies to achieve network size compression while balancing performance. Simulations on classical chaotic systems demonstrate that the optimized network retains only a subset of core neurons from the original network while achieving a comparable predictive performance. Comparative experiments further reveal the differential effects of various pruning strategies and confirm that the hybrid regularization method outperforms other approaches in balancing network compactness and stability, and shows stronger universality. Overall, this work provides an efficient solution for dynamic system modeling in resource-constrained scenarios and physical realization.