Abstract
Artificial Immune Networks (AIN) is a population-based evolutionary algorithm that is inspired by theoretical immunology. It applies ideas and metaphors from the biological immune system to solve multi-disciplinary problems. This paper presents a novel application of the AIN for optimizing shallow machine learning (ML) classification algorithms. AIN accomplishes this task by searching the best hyper-parameter set for a specific classification algorithm (also termed model selection), which minimizes training error and enhances the generalization capability of the algorithm. We present a convergence analysis of the proposed algorithm and employ it in conjunction with selected, well-known ML classifiers, namely, an extreme learning machine (ELM), a support vector machine (SVM) and an echo state network (ESN). The performance is evaluated in terms of classification accuracy and learning time, using a range of benchmark datasets, and compared against grid search as well as evolutionary strategy (ES)-based optimization techniques. An empirical study with different datasets demonstrates improved classification accuracy of SVM, from 2% to 5%, for ESN from 3% to 6%, whereas in the case of ELM from 3% to 9%. Comparative simulation results demonstrate the potential of AIN as an alternative optimizer for shallow ML algorithms.