Cost Effective Computationally Intelligent Control – An Augmented Switching Manifold Approach Mehmet Önder Efe Okyay Kaynak Carnegie Mellon University Bogazici University Electrical and Computer Engineering Department Electrical and Electronic Engineering Department Pittsburgh, PA 15213-3890 Bebek, 80815, Istanbul U.S.A. TURKEY efemond@andrew.cmu.edu kaynak@boun.edu.tr Xinghuo Yu Serdar Iplikçi Central Queensland University Bogazici University Faculty of Informatics and Communication Electrical and Electronic Engineering Department Rockhampton, QLD 4702 Bebek, 80815, Istanbul AUSTRALIA TURKEY X.Yu@cqu.edu.au iplikcis@boun.edu.tr Abstract - In this paper, a novel method for tuning the parameters of a class of intelligent PD controllers is discussed. The aim of the design is to extract a tuning law such that the specifications of the control problem are met and the adjustable parameters evolve bounded. The achievement of these specifications is a challenge in the presence of strong external disturbances, ambiguities in the plant model and nonlinearities, which absolutely require robustness for performance and stability for safety and applicability. The approach introduced in this paper achieves these targets by utilizing the sliding mode control technique based on an augmented switching manifold. The proposed law is applicable to the class of controllers, the output of each member of which is linear in the adjustable parameter set. This stipulates that the application spectrum of the algorithm extends from PID controller to fuzzy controllers and some structures of neural controllers. In the application example, control of a coupled double pendulum system is considered. The dynamic model of the plant is assumed to be unknown and the difficulties introduced by observation noise are studied. I. INTRODUCTION The problem of tuning the parameters of a controller for meeting a set of predefined performance specifications is a challenge because of the nonlinearities existing in the plant model, disturbances and time varying nature of the processes. Especially if the accuracy in the response is sought, the controller must have a degree of robustified intelligence so that the nonlinear behavior is handled together with disturbance rejection ability. One must now question how the designer can define robustified intelligence and achieve it with the known design tools. The concept of intelligence in this context should refer to the acquisition of the current state of the system under investigation and generating an appropriate decision with an increased autonomy. In this respect, the design of a training strategy necessitates the separation of useful knowledge and disturbance related components contained in the training signals, which directly influence the evolution of the parameters and consequently the output of the intelligent system. Robustified intelligence accounts for the behavior in the parameter space, the motions in which are characterized by the adopted training strategy; and the robustness in this space can be defined as the occurrence of a parametric evolution in finite volume. The studies reporting the use of Sliding Mode Control (SMC) for parameter tuning in Computational Intelligence (CI) by Sanner and Slotine [1], and Sira-Ramirez and Colina- Morles [2] have been the stimulants, which proved that the robustness feature of SMC could be exploited in the training of the architectures of CI. These studies pioneered a vast majority of researchers working on SMC and CI. Sanner and Slotine considered the training of GRBFNN which has certain degrees of analytical tractability in explaining the stability issues, and Sira-Ramirez et al have shown the use of ADALINEs with a SMC based learning strategy. As an illustrative example, the inverse dynamics identification of a Kapitsa pendulum has been demonstrated together with a thorough analysis towards the handling of disturbances. Hsu and Real [3-4] demonstrate the use of SMC with Gaussian NNs, Yu et al [5] introduces the dynamic uncertainty adaptation of what is proposed in [2], and demonstrate the performance of the scheme on the Kapitsa pendulum. Parma et al [6] use the SMC technique in parameter tuning process of multilayer perceptron. Latest studies towards the integration of SMC and CI have shown that the tuning can be implemented in dynamic weight filter neurons [7] and in parameters of a controller [8]. A different viewpoint towards this integration is due to Efe et al [9-10], which has the goal of reducing the adverse effects of noise driven parameter tuning activity in gradient techniques. The key idea in these works is the mix two training signals in a weighted average sense. A good deal of review is provided in the recent survey of Kaynak et al [11]. The survey illustrates how SMC can be used for training in CI as well as how CI can be utilized for the tuning of parameters in conventional SMC. In [8,12], it is presented that the original form of the method discussed by Ramirez et al [2] can be used for control applications, in which the target output of the intelligent system, i.e. the controller, is unknown. The major difference of what is presented in this paper from what has been discussed in the literature is the construction of a IECON'01: The 27th Annual Conference of the IEEE Industrial Electronics Society 0-7803-7108-9/01/$10.00 (C)2001 IEEE 31