A.J. Krijgsman

Artificial intelligence in real-time control

Abstract This review paper suggests that established control techniques are incapable of achieving the last few percentage of efficiency needed in complex, fully-integrated industrial automation. Thus, we have to look towards new philosophies, and an investigation into techniques emerging from the world of artificial intelligence seems to be appropriate. Expert systems, fuzzy logic and neural networks are amongst the first tools to offer potential. However, because most current AI applications tend to be in non-real-time areas, the question of the use of AI in hard real-time situations must be examined, as must the question of distributed AI. It is concluded that the key to future control systems lies in real-time distributed AI systems.

Predictive control based on neural network models with I/O feedback linearization

This paper presents an approach for the constrained non-linear predictive control problem based on the input-output feedback linearization (IOFL) of a general non-linear system modelled by a discrete-time affine neural network model. Using the resulting linear system in the formulation of the original non-linear predictive control problem enables to restate the optimization problem as the minimization of a quadratic function, which solution can be found using reliable and fast quadratic programming (QP) routines. However, the presence of a non-linear feedback linearizing controller maps the original linear input constraints onto non-linear and state dependent constraints on the controller output, which invalidates a direct application of QP routines. In order to cope with this problem and still be able to use QP routines, an approximate method is proposed which simultaneously guarantees a feasible solution without constraints violation over the complete prediction horizon within a finite number of steps, ...

Constrained Nonlinear Predictive Control Based on Input-Output Linearization Using a Neural Network

Abstract Affine neural network models can be used as good aproximators of the dynamics of a nonlinear process, and are easily included in a input-output feedback linearization (IOFL) scheme. This paper proposes a new solution for solving a constrained optimization problem using IOFL imbedded in a predictive control scheme. The linearization of the nonlinear feedback law over the entire prediction horizon, enables an optimal solution to be found by solving a general quadratic programming problem. The procedure here presented also guarantees convergenceto a feasible solution without constraint violation.

A new method for identification and control of nonlinear dynamic systems

Abstract This paper presents a new model for nonlinear systems which consists of two parts: a linear part and a static nonlinear output part. The linear part is a linear combination of the models outputs, where the static nonlinear function maps the output of the linear part to the models output. This model can be applied to represent a relatively large class of nonlinear dynamic systems with fading memory. Based on this model, one can directly introduce the corresponding output feedback control law. The identification and control method is applied to two simulation examples of a discrete-time system and the dynamics of a complicated missile to demonstrate the performance and efficiency of the proposed method.

System Identification with Orthogonal Basis Functions and Neural Networks

For the control of a process, usually the relation between past input-output data of the process and future outputs must be identified. For the identification of nonlinear systems, neural networks can be used [3]. In this context, neural networks are nonlinear black-box models, to be used with convential parameter estimation methods. Two important models are: NNFIR-models: Neural Network Finite Impulse Response models, which use only past process inputs u(k − n) as inputs for the network; NNARX-models: Neural Network Auto Regressive with eXogeneous input models, which use past process inputs u(k − n) and past process outputs y(k − n) as inputs for the network.

Dice: A Real-Time Toolbox

Abstract In this paper research the DICE toolbox is described. This toolbox has been developed for real-time AI issues. DICE offers progressive reasoning as a way to handle time critical situations. DICE is a multitasking environment developed for a VAX/VMS environment. It offers both boolean and multivalued logic implementations. Some examples concerning the syntax are given.

Single-layer networks for nonlinear system identification

Abstract The modelling of unknown nonlinear system dynamics is a useful activity for model-based nonlinear control. Single-layer network models, a special class of neural networks, make interesting, practical modelling of nonlinear system dynamics possible. The optimization and analysis of single-layer network models are tractable problems, since single-layer networks constitute linear regression models. Several model representations can be cast into single-layer network form, like Wiener models and parametric approximating functions. Wiener models suffer from the disadvantage of having large numbers of parameters, while approximating functions generally allow a quite modest model size. Especially, certain radial basis type functions (RBF) for function approximation not only have effective modelling capacities, but provide stable models as well. The application of combined linear/RBF models appears to improve model optimization and quality significantly. The orthogonal least squares selection algorithm has proved to be an effective and reliable off-line optimization tool for single-layer network-type models of any representation, but it suffers from high calculation loads. Examples are given, demonstrating the capabilities of the described methods.

Dice: A framework for intelligent real-time control

Abstract This article presents the development of a real-time intelligent control environment: DICE. Main emphasis is given to the requirements of knowledge-based systems to be realized in a real-time system. This results in a multi-task environment with several expert system kernels running in parallel. A blackboard mechanism is used for intertask communication. A kernel of the system uses demons to control other expert system kernels. DICE offers both boolean and multiple valued logic reasoning in separate expert system modules. As a worked out example an adaptive fuzzy logic based controller is described. Finally attention is paid to a general intelligent control scheme. Shown is that DICE is very well suited for the tasks involved in such a strategy. The role of neural networks in this configuration is also treated. The article ends with some conclusions about the use of DICE as a framework for intelligent control.

Knowledge-Based Tuning and Control

Abstract This article presents an in-line knowledge-based supervisory and tuning system for predictive control. The set up for the system is discussed and a summary is given of the control algorithm and the parameters to be tuned and initialized. After an initialisation phase in which certain parameters are set in accordance to the data required by the system, fine-tuning is performed in-line and the control system is brought to a state in which the control requirements are fulfilled. The results of the system, tuning a predictive controller, are shown in a real-time experiment.

Rule-based controller using fuzzy logic

This paper describes research into direct real-time fuzzy expert control. In order to carry out this research a fuzzy expert system shell was developed. As an application a fuzzy rule-based controller was designed. The implemented rule base uses two control strategies:direct expert control (d.e.c.) and direct reference expert control (d.r.e.c.). Simulation results show that a wide range of processes can be controlled with little a priori information about the process dynamics. Adaptation of the rule-based controller is used to adapt the rule-based controller to the specific process to be controlled, resulting in a better control performance.