Neural network-based adaptive

Having successfully implemented conventional MRAC techniques, the next logical step was to try to incorporate the MRAC techniques into a neural network-based adaptive control system. The ability of multilayered neural networks to approximate linear as well as nonlinear functions is well documented and has foimd extensive application in the area of system identification and adaptive control. The noise-rejection properties of neural networks makes them particularly useful in smart structure applications. Adaptive control schemes require only limited a priori knowledge about the system to be controlled. The methodology also involves identification of the plant model, followed by adaptation of the controller parameters based on a continuously updated plant model. These properties of adaptive control methods makes neural networks ideally suited for both identification and control aspects [7-11]. 

A schematic diagram of the neural network-based adaptive control technique is shown in Fig. 4.9. A neural network identification model is trained using a static backpropagation algorithm to generate p(fc + 1), given past values of y and u. The identification error is then used to update the weights of the neural identification model. The control error is used to update the... 

Various neural network-based adaptive control techniques were discussed in this study. A major problem in implementing neural network-based MRACs is the translation of the output error between the plant and the reference model so as to train the neural controller. A technique called iterative inversion, which inverts the neural identification model of the plant for calculating neural controller gains, has been used. Due to the real-time computer hardware limitations, the performance of neural network-based adaptive control systems is verified using simulation studies only. These results show that neural-network based MRACs can be designed and implemented on smart structures. 

The end users of CBR systems should in principle be able to maintain the case-bases themselves and use the systems for varying inspection types (within certain limits). Adaptation of neural-network based systems, though possible by end-users, is difficult to be done reliably. Adaptation of rule-based systems usually has to be done by the rule-base designer. 

Neural Network-Based Model Reference Adaptive Control... 

In the nemal network-based adaptive control scheme, a neurocontroller is trained to approximate an inverse model of the plant. We have introduced an adaptive activation function for increasing the training rate of the neural controller, and the proposed function is described in this section. 

A neural network-based model reference adaptive control scheme for nonlinear plants is presented in this section. 

Neural Network-Based Optimizing Controller With On-Line Adaptation... 

In this section, a neural network-based design methodology is developed that utilizes the adaptability of neural networks to compensate for the time varying dynamical properties of smart structures. This formulation is designed to be implemented using the ETANN chip and also allows the designer to directly incorporate all the a priori information about the system that may be available. An important feature of this formulation is that it relies only on the experimental input/output data of the system for the design. The ability of neural networks to map nonlinear systems allows this formulation to be extended to incorporate nonlinearity in structural systems. 

Hoskins, D.A. Neural Network Based Model-Reference Adaptive Control. Ph. D. Dissertation, University of Washington, UMI Dissertation Services, Ann Arbor, MI (1990)... 

Polymer electronics, including the use of polymer-nanocomposite-based devices, provide a number of alternative approaches, such as the use of adaptive circuits or the neural network-based processor architectures. Combined with a better understanding of the conductivity mechanism in conjugated polymers such as poly(acetylene) [2, 3] and poly(thiophenes) [4], these factors have initiated this second wave of interest in the low molecular weight organic and polymer-based optoelectronic, electronic and photonic devices. 

The control law This is the information flow structure through which the manipulated variables are handled based on the measurements. The complexity of the control law is determined by the diversity of the control objective. As a result, the controller can be simple (on—off, proportional, proportional-integrated differential), more complicated adaptive model-based, empirical (expert systems), fuzzy or neural network-based. Detailed references on the various control systems applied on anaerobic digesters can be found in Boe (2006) and Find et al. (2003). 

Eor a number of cognitive or interpretive tasks, there are alternatives to mainstream knowledge-based systems that may be more appropriate, especially if adaptive behavior and learning capabihty are important to system performance. Two approaches that embody these characteristics are neural networks (nets) and case-based reasoning. 

Numeric-to-numeric transformations are used as empirical mathematical models where the adaptive characteristics of neural networks learn to map between numeric sets of input-output data. In these modehng apphcations, neural networks are used as an alternative to traditional data regression schemes based on regression of plant data. Backpropagation networks have been widely used for this purpose. 

Nonmodel-based controllers, such as the least mean square (LMS) and artificial neural network back-propagation adaptive controllers, employ iterative approaches to update control parameters in real time [14-17]. However, those methods may encounter difficulties of numerical divergence and local optimiza-... 

An adaptation of the simple feed-forward network that has been used successfully to model time dependencies is the so-called recurrent neural network. Here, an additional layer (referred to as the context layer) is added. In effect, this means that there is an additional connection from the hidden layer neuron to itself. Each time a data pattern is presented to the network, the neuron computes its output function just as it does in a simple MLP. However, its input now contains a term that reflects the state of the network before the data pattern was seen. Therefore, for subsequent data patterns, the hidden and output nodes will depend on everything the network has seen so far. For recurrent neural networks, therefore, the network behaviour is based on its history. 

In this work, new developments were achieved through the use of new examples, one of which the optimisation of a real crude distillation unit involving 19 decision variables. The performance of the metamodel-based optimisation is compared with results obtained with the optimisation based on a first-principles model, embedded in a sequential-modular process simulator. It is shown that metamodel-based optimisation with adaptation of the metamodels during the optimisation procedure provides results with good accuracy and significant reduction of computational effort. The performance comparison between neural networks and kriging models for chemical processes is another contribution of this work. 

---