Methaheuristics (MHs) are techniques widely used for solving complex optimization problems. In recent years, the interest in combining MH and machine learning (ML) has grown. This integration can occur mainly in two ways: ML-in-MH and MH-in-ML. In the present work, we combine the techniques in both ways—ML-in-MH-in-ML, providing an approach in which ML is considered to improve the performance of an evolutionary algorithm (EA), whose solutions encode parameters of an ML model—artificial neural network (ANN). Our approach called TS \(_{in}\) EA \(_{in}\) ANN employs a reinforcement learning neighborhood (RLN) mutation based on Thompson sampling (TS). TS is a parameterless reinforcement learning method, used here to boost the EA performance. In the experiments, every candidate ANN solves a regression problem known as protein structure prediction deviation. We consider two protein datasets, one with 16,382 and the other with 45,730 samples. The results show that TS \(_{in}\) EA \(_{in}\) ANN performs significantly better than a canonical genetic algorithm (GA \(_{in}\) ANN) and the evolutionary algorithm without reinforcement learning (EA \(_{in}\) ANN). Analyses of the parameter’s frequency are also performed comparing the approaches. Finally, comparisons with the literature show that except for one particular case in the largest dataset, TS \(_{in}\) EA \(_{in}\) ANN outperforms other approaches considered the state of the art for the addressed datasets.
