Copyright © IFAC Artificial Intelligence in Real-Time Control, Delft, The Netherlands, 1992 GENETIC ALGORITHMS FOR PROCESS CONTROL: A SURVEY J.M. Renders*, J.P. Nordvlk* and H. Berslni** "Commission of the European Communities, Joint Research Centre, Institute/or Systems Engineering and Informatics, 1-21020 Ispra (Va), Italy **lRlDIA - Universite Libre de Bruxelles CP 149/6, 50 avoFr. Roosevelt, 8-1050 Bruxelles, Belgium Abstract. This paper presents a survey of the potential use of Genetic Algorithms (GAs) for process control. GAs are a family of iterative search algorithms based on an analogy with the process of natural selection and evolutionary genetic. Application to off-line control is first en- visaged, where GAs are used for task scheduling, calculation of optimal set points and design of optimal control strategies. Then application to on-line control is considered, focusing on system identification and enhancement of existing controllers, two problems for which GAs seem to of- fer the most promising results. After the description of possible applications of GAs to supervi- sory problem, the general advantages, drawbacks and limitations of applying GAs to process control are discussed, and further lines of research are drawn. Keywords. Adaptive Systems, Control System Design, Learning Systems, System Identification, Genetic Algorithms. INTRODUCTION Recent interest in paradigms inspired by biological metaphors for adaptive process control stems mainly from the difficulties of traditional control theory to deal with complex varying or uncertain environment. These difficulties mainly lie in the fact that even traditional adaptive control theory requires precise knowledge of the process to be controlled (slow deviation of the proc- ess parameters with respect to the adaptive capacity of the controller, a priori knowledge about the structure of the process and its perturbations) [Astrom and Witten- mark, 1989J. Moreover this knowledge has to be precise and complete: any slight imprecision can degrade dra- matically the quality of the control. On the other hand, several biological adaptive systems, whose principles and mechanisms have inspired artificial computing metaphors such as Neural Networks (NN) [Rumelhardt and Mc Clelland, 1986], Immune Networks (IN) [Atlan and Cohen, 1989; Varela, Sanchez and Coutinho, 1989] and Genetic Algorithms (GAs) [Goldberg, 1989; Davis 1991] are characterized by their ability to reconfigure themselves to an unspecified environment in an incre- mental and robust fashion. These adaptive biological systems seem to have less stringent requirements than the traditional control theory - or at least different ones. They therefore constitute a promising approach for proc- ess control, although these biological metaphors are still under development and only partial applications are available. While there already exists an abundant literature on ar- tificial neural networks applied to process control [Miller, Sutton and Werbos, 1990; Narendra and Parthasarathy, 1990], the use of the IN and GA tech- 323 niques for process control is more recent and not so de- veloped [Varela and Bersini, 1991; De long, 1980 and 1988J. This paper focuses on the GA technique and pre- sents a survey of its applications to process control, em- phasizing the originality, advantages and drawbacks of the various approaches, and attempts to determine the most promising lines of research. The paper is organized as follows. First a brief descrip- tion of GAs is given. Then applications of GAs to off- line control, on-line control and supervisory control are described successively. Finally conclusions about the use of GAs in process control are presented as well as some guidelines for further investigations in this do- main. BRIEF DESCRIPTION OF GAS Genetic Algorithms are a family of iterative search algo- rithms based on an analogy with the process of natural selection (Darwinism) and evolutionary genetics. The search aims to optimize a user-defined function called the fitness function. To perform this task, GAs maintain a "population" of candidate points in the search space, called "individuals". During each iteration, called a "generation", anew population is created. This new gen- eration generally consists of individuals that fit better than the previous ones to the external environment as represented by the fitness function. As the population it- erates through successive generations, the individuals will in general tend towards the optimum of the fitness function. To generate a new population on the basis of a previous one, GAs perform three steps: (a) they evaluate the fitness score of each individual of the old population, (b) they select individuals on the basis of their fitness score, (c) they recombine these selected individuals us-