Neuroevolutive Control of Industrial Processes Through Mapping Elites

Abstract

Classical model-based control techniques used in process control applications present a trade-off between performance and computational load, specially when using complex nonlinear methods. Learning-based techniques that allow the controller to learn policies from data represent an appealing alternative with potential to reduce the computational burden of real-time optimization. This work presents an efficient learning-based neural controller, optimized using evolutionary algorithms, designed specially for maintaining diversity of individuals. The search of solutions is conducted in the parameters space of a population of deep neural networks, which are efficiently encoded with a novel compression algorithm. Evaluation against strong baselines demonstrates that the proposed controller achieves better performance in most of the chosen evaluation metrics. Results suggest that learning-based controllers are a promising option for next generation process control in the context of industry 4.0.