Adapting the rule-base

Abstract

A rule-based system for optimizing combustion in multiple-burner installations was built and tested on the furnace of a continuous annealing line for rolled steel. The furnace has only one firing level, and the rules were elicited from energy experts with this problem in mind. The system was then installed on a multiple-burner boiler in the steel industry, but it did not respond very fast to changes in the firing level of the boiler. One particular solution to this problem is to enter new rules into the rule base to deal with the changed situation. A more general solution is to build into the rule base a learning component to help it to cope with new situations. Promising results have been obtained with the genetic algorithm, which has proved to be robust in this noisy domain and suitable for learning control rules that give performance comparable to that of rules elicited from the experts. Experiments carried out on simulations of multiple-burner installations with two firing levels, using the genetic algorithm to learn the best actions for given situations, are described, and the results are discussed.<>