A learning technique for a general purpose optimizer

The goal of the machine learning method implemented in this article is to broaden the region of operability of an adaptive control system by switching multiple controller models. The learning system determines a separate set of control parameter values, for optimal performance under given operating conditions, and stores them in memory. In this way, the controller is able to operate effectively over the whole environment. The basic scheme implements a single neuron, the perceptron, which approximates the process model and then directly computes the control signals. An example application is also described of an innovative sensing method, which has been developed to replace leaf sensors in plant propagation chambers, by emulating the sensor in software. Such chambers present critical situations for control because of the high humidity levels required, which makes direct sensing methods unsuitable. The proposed method enhanced the reliability of the control system and eliminated the need for costly electronic leaf sensors and the associated need for great care and frequent calibration. The method in principle combines ordinary measurements of ambient temperature, humidity and radiation, to calculate the controls of the humidification process in mist or fog propagation chambers. The performance surface was studied and a modification of the searching algorithm has improved the learning rate significantly. The method is applicable to any system whose performance can be defined and measured by simulation or experiment. (C) 2000 Elsevier Science B.V.