Neural Networks for Bespoke Control

Abstract Artificial intelligence techniques are starting to suggest some intriguing paradigm changes for real-time control. We discuss one novel application of neurocontrol in this context: the realization of made-to-order, or bespoke, controllers. With an appropriate optimization technique, neural-network-based nonlinear controllers can be designed to accommodate arbitrary application-specific requirements. The design requires extensive offline optimization: CPU days for single-loop problems. The on-line cost is minimal. Some simulation experiments are presented that adopt practically relevant definitions of closed-loop response features and controller robustness—definitions that cannot be accommodated with conventional control technology. We also outline the non-gradient-based optimization algorithm that we have developed in this research.