Neural networks and fuzzy logic made simple

Engineers and scientists are in a constant strive to emulate the learning process in the human brain, it's ability to generalize and adapt to changes. The neural network technology was developed with this goal in mind, and is currently being used in real time control applications benefiting from these advanced features. Fuzzy logic on the other hand is a problem solving technique that uses approximate description of the system using simple rules in a human language like English. At first glance, these technologies appear difficult to comprehend and intimidating to use in an actual application. This paper will bring both of these technologies to an understanding level and will show how the combination of both technologies, NeuFuz, takes advantage of both Neural Networks and Fuzzy Logic by eliminating their drawbacks, thus providing the design engineer with a very simple tool that reduces design time and provides a more effective solution. A design of a fan controller will be discussed as an application of this technology. 1.0 Introduction Control applications today demand more and more intelligence and flexibility. These requirements demand for a control system that can handle non linearity and time variation. Moreover the control system should be simple to understand and easy to develop. Conventional techniques are not well suited for nonlinear time variant applications. Neural networks is an elegant solution for nonlinear applications. It requires a number of data sets representing the output as a function of the inputs, covering the entire range of operation. This data is used to train the neural network. Once the training is completed the neural network can generalize and determine an output from any input applied. The neural network is a network of interconnected neurons, where each neuron has the ability to compute an output signal based on all the incoming signals from the neurons connected to it in the previous stage. This output signal is then multiplied by the corresponding weights (the connection strength) before being transmitted to the neuron on the next stage. The neural network weights, or in other words, the connection strengths between neurons are determined in a process called learning, ISBN# 0-7803-2636-9 596 the process of training the neural network for a specific application to a desired level of accuracy. Although this technology can handle non linear system requirements well but is often expensive to implement. Fuzzy Logic is another technique that has been proven very successhl in solving nonlinear problems in many areas where mathematical modeling of the system is difficult or costly to implement. Fuzzy logic offers in these design cases simplicity and ease in design. However Fuzzy logic has some problems as well: The process in Fuzzy Logic is to determine membership functions for the inputs and outputs and generate a number of rules for the application . One important factor here is that membership functions and rules have a substantial effect on the output and therefor on the operation of the control system. When the complexity of the system increases, the set of membership fknctions and rules are difficult to generate. Moreover, once a set of membership functions and rules is determined for an application, it is necessary to properly tune them for a better solution. This is a time consuming task that can take weeks to months in some complex systems. Both of these technologies have some shortcomings, but the combination of the two offers an excellent solution that is made simple, and is based on the advantages of both, alleviating the drawbacks. In addition, this solution automatically generates code that implements the application in either COP800 specific assembly code or in ANSI C language. 2.0 The Combined Neural network Fuzzy Logic Solution (NeuFuz): As discussed in the introduction, both technologies neural networks and kzzy logic have many advantages and some disadvantages. The combined solution takes advantage of the neural networks learning capability and applies it to resolve the membership functions and rules generation of the Fuzzy logic, therefor coming up with an integrated solution that offers a tremendous tool to the design engineer. The most significant advantage of neural networks, is the ability of the net to learn the system behavior of the application. A process in which the engineer applies a number of data points representing the output as a function of the inputs selected in the control application. The data points used to train the net are a result of the designer’s knowledge of the system’s expected fbnctionality and determined by the system requirements, measurements or simulations. The data points are also a partial representation of the control surface of the system. The neural network has the ability to generalize and determine the output for any input value applied. The training process of the neural network is applying the input values to the net and adjusting the weights of the neuron connections between the layers to get the appropriate output value to the required accuracy of the system. The trained network and internal weight distribution can now be mapped into membership fimctions and kzzy rules. As can be seen in this discussion, the neural network advantage is used to resolve the above mentioned problem associated with hzzy logic rules and membership functions. The integration of both of these technologies