Page 253 - Autonomous Mobile Robots
P. 253

Adaptive Neural-Fuzzy Control of Mobile Robots             239

                              For clarity, let us define the weight vector and fuzzy basis function vector
                              respectively as
                                                                        T
                                                                        ]
                                                      W =[w 1 , w 2 , ... , w n r
                                                                       T
                                                                      ]
                                                S(x, c, σ) =[s 1 , s 2 , ... , s n r
                                          n i         n r  n i              T  T     T T

                              where s l =    µ l(x i )/[      µ k(x i )], c =[c , c , ... , c ] , and
                                          i=1  A i    k=1  i=1  A i         1  2     n r
                                              T T
                                       T
                                    T
                              σ =[σ , σ , ... , σ ] . Then, Equation (6.13) can be represented as
                                   1   2     n r
                                                           T
                                                      y = W S(x, c, σ)                 (6.15)
                              Remark 6.2  For Equation (6.15), W and S(x, c, σ) are the weights and the
                              (normalized) basis functions in NN terminology, while they are the outputs of
                              the rules and the weighted firing strength in fuzzy logic terminology. Because
                              of the difficulty in deriving the rules in fuzzy systems for systems with little
                              physical insights, we would hereby like to present the adaptive laws to design
                              the outputs of the “rules” numerically using adaptive (NN) control techniques.
                                 It has been proven that, if the number of the fuzzy rules n r is sufficiently
                              large, a fuzzy logic system (6.15) is capable of uniformly approximating any
                              given real continuous function, h(x), over a compact set   x ⊂ R n i  to any
                              arbitrary degree of accuracy in the form
                                                  T
                                                 ∗
                                                       ∗
                                                          ∗
                                         h(x) = W S(x, c , σ ) +  (x),  ∀x ∈   x ⊂ R n i  (6.16)
                                         ∗
                                      ∗
                              where W , c , and σ  ∗  are the ideal constant vectors, and  (x) is the
                                                                                   ∗
                                                                                       ∗
                                                                                           ∗
                              approximation error. The following assumption is made for W , c , σ ,
                              and  (x).
                                                                  ∗
                                                                ∗
                                                                     ∗
                              Assumption 6.2  The ideal NF vectors W ,c , σ , and the NF approximation
                              error are bounded over the compact set, that is,
                                                                 ∗
                                                     ∗
                                         ∗
                                       W  ≤ w m ,   c  ≤ c m ,   σ  ≤ σ m ,  | (x)|≤   ∗
                              ∀x ∈   x with w m ,c m , σ m , and   being unknown positive constants.
                                                       ∗
                                                                   ∗
                                                                      ∗
                              Remark 6.3  The optimal weight vector W ,c , and σ  ∗  in (6.16) is an
                                                                                           ∗
                                                                                        ∗
                              “artificial” quantity required only for analytical purposes. Typically, W ,c ,
                                   ∗
                                                                                          n i
                              and σ are chosen as the value of W that minimizes  (x) for all x ∈   x ⊂ R ,
                              that is,

                                                                         T
                                         ∗
                                               ∗
                                            ∗
                                       (W , c , σ ) := arg min  sup |h(x) − W S(x, c, σ)|
                                                       W,c,σ
                                                            x∈  x
                              © 2006 by Taylor & Francis Group, LLC
                                 FRANKL: “dk6033_c006” — 2006/3/31 — 16:42 — page 239 — #11
   248   249   250   251   252   253   254   255   256   257   258