Combining neutral networks and control: potentialities, paterns and perspectives

Machine learning tools are widely used for knowledge extraction, modeling, and decision tasks; a range of problems that Control Theory also tackles. Their relations have been largely explored by looking at stochastic control and Markov Decision Processes, due to the proximity of their formulations. However, novel links between machine learning and deterministic control are emerging; combining both approaches, e.g. by performing identification with learning, or controlling the training process. The recent flourishing literature is vast: there is a need to identify challenges, trends and opportunities on this interface. This survey contributes i) to the compared analysis of both fields. ii) Based on literature review, a categorization of combinations of learning and control is drawn. In the control framework, learning has been used for modeling, controllers tuning or adaptation, generating a controller or as a controller itself, for translating complex objectives, or checking controlled systems. Conversely, in the learning framework, control is used for tuning hyperparameters, selecting or generating training data, as the training or decision-making algorithm itself or to guarantee learning properties. iii) Finally, discussions on the literature open novel promising combinations to be explored, such as control of neural networks' training process.