This set of proceedings contains the most significant papers presented at the third IFAC Workshop on Artificial Intelligence in Real-time Control, which was held from September 23-25, 1991 in the USA. In this workshop, although there were still some "exotic" applications, a more practical view of the applications and limitations of current AI technology dominated the participants' discussions. With its resultant focus on reliability and safety considerations, the workshop posed as many questions as it answered. It provides an excellent mirror of the current state-of-the-art which these proceedings are intended to illustrate.

The symposium had two main aims, to investigate the state-of-the-art in the application of artificial intelligence techniques in real-time control, and to bring together control system specialists, artificial intelligence specialists and end-users. Many professional engineers working in industry feel that the gap between theory and practice in applying control and systems theory is widening, despite efforts to develop control algorithms. Papers presented at the meeting ranged from the theoretical aspects to the practical applications of artificial intelligence in real-time control. Themes were: the methodology of artificial intelligence techniques in control engineering; the application of artificial intelligence techniques in different areas of control; and hardware and software requirements. This symposium showed that there exist alternative possibilities for control based on artificial intelligence techniques.

Artificial Intelligence is one of the new technologies that has contributed to the successful development and implementation of powerful and friendly control systems. These systems are more attractive to end-users shortening the gap between control theory applications. The IFAC Symposia on Artificial Intelligence in Real Time Control provides the forum to exchange ideas and results among the leading researchers and practitioners in the field. This publication brings together the papers presented at the latest in the series and provides a key evaluation of present and future developments of Artificial Intelligence in Real Time Control system technologies.

Computer scientists have long appreciated that the relationship between algorithms and architecture is crucial. Broadly speaking the more specialized the architecture is to a particular algorithm then the more efficient will be the computation. The penalty is that the architecture will become useless for computing anything other than that algorithm. This message holds for the algorithms used in real-time automatic control as much as any other field. These Proceedings will provide researchers in this field with a useful up-to-date reference source of recent developments.

Papers presented at the workshop are representative of the state-of-the art of artificial intelligence in real-time control. The issues covered included the use of AI methods in the design, implementation, testing, maintenance and operation of real-time control systems. While the focus was on the fundamental aspects of the methodologies and technologies, there were some applications papers which helped to put emerging theories into perspective. The four main subjects were architectural issues; knowledge - acquisition and learning; techniques; and scheduling, monitoring and management.

Proceedings of the European Control Conference 1991, July 2-5, 1991, Grenoble, France

This book focuses on how the BOXES Methodology, which is based on the work of Donald Michie, is applied to ill-defined real-time control systems with minimal a priori knowledge of the system. The method is applied to a variety of systems including the familiar pole and cart. This second edition includes a new section that covers some further observations and thoughts, problems, and evolutionary extensions that the reader will find useful in their own implementation of the method. This second edition includes a new section on how to handle jittering about a system boundary which in turn causes replicated run times to become part of the learning mechanism. It also addresses the aging of data values using a forgetfulness factor that causes wrong values of merit to be calculated. Another question that is addressed is “Should a BOXES cell ever be considered fully trained and, if so, excluded from further dynamic updates”. Finally, it expands on how system boundaries may be shifted using data from many runs using an evolutionary paradigm.

Robust control mechanisms customarily require knowledge of the system’s describing equations which may be of the high order differential type. In order to produce these equations, mathematical models can often be derived and correlated with measured dynamic behavior. There are two flaws in this approach one is the level of inexactness introduced by linearizations and the other when no model is apparent. Several years ago a new genre of control systems came to light that are much less dependent on differential models such as fuzzy logic and genetic algorithms. Both of these soft computing solutions require quite considerable a priori system knowledge to create a control scheme and sometimes complicated training program before they can be implemented in a real world dynamic system. Michie and Chambers’ BOXES methodology created a black box system that was designed to control a mechanically unstable system with very little a priori system knowledge, linearization or approximation. All the method needed was some notion of maximum and minimum values for the state variables and a set of boundaries that divided each variable into an integer state number. The BOXES Methodology applies the method to a variety of systems including continuous and chaotic dynamic systems, and discusses how it may be possible to create a generic control method that is self organizing and adaptive that learns with the assistance of near neighbouring states. The BOXES Methodology introduces students at the undergraduate and master’s level to black box dynamic system control , and gives lecturers access to background materials that can be used in their courses in support of student research and classroom presentations in novel control systems and real-time applications of artificial intelligence. Designers are provided with a novel method of optimization and controller design when the equations of a system are difficult or unknown. Researchers interested in artificial intelligence (AI) research and models of the brain and practitioners from other areas of biology and technology are given an insight into how AI software can be written and adapted to operate in real-time.

The 47 papers in this volume provide a useful reference tool for the state-of-the-art research in real-time programming.