differentiable functions of these parameters. 13.2.2 Adaptive Neuro-fuzzy Inference System (ANFIS) y inference syste, quite often referred to as the adaptive network based asa The adaptive neuro-fu inference system (ANPIS) is a class of adaptive networks that can function in the 93 inference system. ANFIS uses a hybrid learning algorithna and represents the Sugeno and Tsukamoto fuzzy model which is described below. (a) Sugeno Fuzzy Model, Developed by Takagi, Sugeno and Kang and popularly known as the Sugeno model, this method generates fuzzy rules from a set of data. A typical rule in the Sugeno Model has the form: Scanned with CamScanner578 | tntvoduction to Soft Computing If (xis A) AND (y is B) Then z= f(x, y) Considering f(s, )) to bea first-order polynomial, we get the fistoorder Sugeno fur (B15 described with the help of the following cules Y model and can Rule 1. I (visA,) and (is B) then f, (= Px tay Rule 2. If (vis A) and (y is 8) then f(t. y= px + ay +7, i - Bay ure. Fig. 13.11 shows the ANF IS architecture for two input first order 5; s including the input and output layers, Bach nn tty yer is explained below. Here O, is the outpen a be " Of the (b) ANFIS architect model with two rules. [t consists of five lay specific job to do, The task performed by every node of the layer &. AYER : LAYER LAYER-2 LAYERS 4 4 LAYERS “i = : Hh Whe fe] Fig. 13.11 ANFIS architecture for 2 input first order Sugeno fuzzy model with 2ruiss Layer 1. Every node in this layer is an adaptive node with a node function as described below. My (x), fori=12 5 =y is (1318) oy {ve Oh foria34 is.a linguistic label associated with this node Bor By Needless to say that the membership hell, triangular, or trapezoidal ris labeled Prod and produces fe Here x (or. y)is the input to node /and A, (or B, ,) fore 0, is the membership grade of a fuzzy set A,, Ay functions can be any parameterized function like the & Layer 2. This layer consists of a number of nodes each of whicl product of all the incoming signals on it as its output Os, =H OX) Hy (y), fori = 12 ‘The output from each of these nodes represents the fire strength of the corresponding rl tively, any other T-norm operator that performs as the AND operator can be use the i node calculates the (aztel Je. Alterna Layer 3, Nodes in layer 3 are fixed nodes labeled Nore, and in this layer, ratio between the ?* rule’ firing strength and the sum of the firing strengths of all the rules. 3.7} 0,28, wy tw! The outputs of these nodes are referred to as the normalized firing strengths. Mee Scanned with CamScannerHybrid Systems | 579 Layer 4. ‘The node function of each node in this layer is presented below. OV, =%, i px+qy +n) (13.18) In expression 13.17, w, is the norm: alized firing strength from layer 3 and {p, q,.r)} isthe parameter set of the * node and is referred to as consequent parameters a Layer 5. This single node layer computes the output of the network by summing all incoming sig- nals. The output is expressed as 052), 7, ee (13.19) ", Training the ANFIS is accomplished by using the hybrid learning algorithm presented by Jang. This algorithm uses least squares method to identify the consequent parameters on the layer 4 during the forward pass, During backward pass the errors are propagated backward and the premise parameters are updated using gradient descent technique. Scanned with CamScannersurcomup suncsion parameter fack-propageense estimation. 4624.2 Constraints of ANFIS fuzzy inference systems anf is is more complex. It is not available for all of hen compared to the general # only supports Sugeno-type systems. Such systems have the following, ihe furzy inference system options and properties: They should be the first- of zeroth-order Sugeno-type systems. output that is obtained using weighted average defuerification, All output 2. They should have a single membership firnctions must be the same type and can be either linear oF constant rules, The number of ourpue membership functions 1m 3. They do not share ust be equal to the number of rules, 4. They must have unity weight for each rule. error would occur. AKO, all the customization IFTIS structure does not comply’ with these constraints then 3 etc allows cannot be accepted by an£is. In simpler words membership ta one’s choice, rather those provided ptions that basic fiz pions and. defuzzificarion functions cannot be, made according, ATLA vt should be used. arngisedt => M Scanned with CamScanner