Extreme learning machine based on Q-learning

Extreme learning machines (ELMs) are prone to numerical instability and overfitting when dealing with ill-conditioned matrices, primarily due to the use of the Moore-Penrose pseudoinverse for output weight estimation. To address these issues, this paper proposes an ELM-based on Q-learning (Q-ELM), in which the determination of output weight is reformulated as a Markov decision process (MDP) by utilizing a reinforcement learning agent to iteratively searches for optimal solution within a discrete action space. A weight clipping mechanism is introduced to impose implicit hard-constrained regularization, effectively suppressing noise amplification. Theoretical analysis establishes the convergence of the proposed algorithm under convex objective functions and derives its noise robustness bounds. Experimental results demonstrate that Q-ELM significantly outperforms the standard ELM, achieving accuracy improvements of up to 5.39 percentage points on the small-sample scenarios (Glass dataset) and 2.19 percentage points on the high-dimensional ill-posed task (SynapseMNIST3D dataset). These gains highlight the enhanced robustness and classification performance of the proposed model. Furthermore, across 19 benchmark datasets and in comparison with nine competing algorithms, Q-ELM attains an average ranking of 4.16, securing the top position among all single-hidden layer feed forward network (SLFN) algorithms. Notably, statistical significance tests confirm that Q-ELM yields superior results compared to conventional ELM and its SLFN variants.