Automatically-Grown Optimal Neural Controller

Designing an optimal controller for practical systems is rewarding yet challenging due to its intricate dynamics, nonlinearity, and time-varying characteristics. This paper proposes an automatic method to design an optimal neural network controller, AutoNN, for practical applications. The AutoNN grows the neural architecture for minimal structural size, with a genetic algorithm to train the discrete structural and continuous weight parameters. This enables the neural controller to generate its required structure adaptively from the distinct characteristics of the plant being controlled, ultimately achieving optimal control. Further, AutoNN alleviates the labour-intensive process of manually adjusting hyperparameters through iterative experimentation. The efficacy of AutoNN is evaluated in designing a nonlinear controller for a complex heating, ventilation, and air conditioning system. The AutoNN approach and its performance are compared with present training methods for neural controllers, such as those based on multilayer perceptron, recurrent, and long short-term memory neural networks. AutoNN achieves a more compact structure with higher control performance than them and the widely used traditional controller, PID.