US5524176A Expired / Lapsed US ⏱️ 0.0 years remaining

Fuzzy expert system learning network

Assignee: Daido Steel Co Ltd

•

Filing: 1994-03-04

•

Granted: 1996-06-04

•

Expires: 2014-03-04

📄 Original PDF 🌐 Google Patents 🔄 Sync Now

Abstract

Abstract In the present invention, prior art techniques are extended to allow application of the backpropagation learning technique to artificial neural networks derived from fuzzy expert system rule-bases. A method in accordance with the invention, referred to herein as a Fuzzy Expert Network (FEN), is implemented in a programmed machine such as a computer to provide automated learning of both "fine" and "coarse" knowledge in a network of artificial neural objects (ANOs) implementing fuzzy modeling rules. Through application of the FEN method, an event-driven fuzzy expert network comprising acyclically connected ANOs derived from fuzzy modelling rules may be implemented. Neural objects implement one or more fuzzy combining and defuzzification rules and use backpropagation of error techniques to implement learning. As in prior art, the FEN allows each ANO to adjust its input weight parameters-"fine" knowledge learning. Unlike prior art, the FEN allows each ANO to modify its internal parameters-"coarse" knowledge learning. This latter action means that individual ANOs have the capability to modify the parameters of the fuzzy rule's membership function upon which they are based. In this way the FEN is able to change the structure of its encoded knowledge over time, making it a more adaptable architecture for autonomous and/or adaptable control systems. Simulation results showing the FEN's learning and adaptability behavior are given.

Inventors

👤 Kazunari Narita
👤 Robert C. Lacher

Bibliographic Details

Application Number —

Jurisdiction US

Data Source google_patents

Last Synchronized 2026-11-03 16:55

CPC Classifications

G; G06; G06N; G06N7/00; G06N7/02; G06N7/04; G06N7/046; G06N3/00; G06N3/02; G06N3/04; G06N3/043

Claims Analysis

Total Claims: 4

Claim 1

Independent Claim 402 words

An acyclic event-driven expert network of artificial neural objects comprising: (a) artificial neural object structures connected by weighted connections; (b) a SUM artificial neural object structure having weighted input signals, x and w, where x i is the value of the i th input and w i is a weight associated with input x i , and internal state y calculated by combining function Γ sum (x), and an output function φ sum (y), wherein (1) the internal state of the SUM artificial neural object structure is generated in accordance with the equation: ##EQU6## and (2) the output of a SUM artificial neural object structure is generated in accordance with the equation: φ.sub.sum (y)=yj (c) a LINEAR THRESHOLD artificial neural object structure having weighted input signals, x and w, where x i is the value of the i th input and w i is a weight associated with input x i , and internal state y calculated by combining function Γ lin (x), and an output function φ lin (y), wherein (1) the internal state of the LINEAR THRESHOLD artificial neural object structure is generated in accordance with the equation: ##EQU7## and (2) the output of a LINEAR THRESHOLD artificial neural object structure is generated in accordance with the equation: ##EQU8## (d) a PRODUCT artificial neural object structure having weighted input signals, x and w, where x i is the value of the i th input and w i is a weight associated with input x i , and internal state y calculated by combining function Γ prod (x), and an output function φ prod (y), wherein (1) the internal state of the PRODUCT artificial neural object structure is generated in accordance with the equation: ##EQU9## and (2) the output of a PRODUCT artificial neural object structure is generated in accordance with the equation: φ.sub.prod (y)=y; and (e) a WEIGHTED-SUM-GRAVITY artificial neural object structure having weighted input signals, x and w, where x i is the value of the i th input and w i is a weight associated with input x i , and internal state y calculated by combining function Γ grav (x), and an output function φ grav (y), wherein (1) the internal state of the WEIGHTED-SUM-GRAVITY artificial neural object structure is generated in accordance with the equation: ##EQU10## and (2) the output of a WEIGHTED-SUM-GRAVITY artificial neural object structure is generated in accordance with the equation: φ.sub.grav (y)=y.

Claim 2

Dependent Claim 402 words

Claim 3

Dependent Claim 286 words

A method for implementing backpropagation learning on the acyclic, event-driven, fuzzy expert network of claim 1, the method comprising the steps of: (a) applying an example vector input signal E to the fuzzy expert network so that a vector output signal z is produced at the output nodes of the fuzzy expert network; (b) generating an vector error signal e for assigning an error value to each node in the fuzzy expert network according to an error equation; (c) generating a new weight vector w new for each node in the fuzzy expert network based on the error value generated in step (b), wherein the new weight vector is generated in accordance with the equation: w.sup.new =w.sup.old +Δw wherein w old is the old weight vector and Δw is a vector pointing in the negative direction of the gradient of the error vector e, wherein the negative gradient of the error vector for a node j is calculated according to: -∇E=2e.sub.j ×φ'(y.sub.j)×z.sub.i ×∂Γ/∂x.sub.ji, wherein e j is the error value associated with node j and wherein z i is node i's output value, wherein x ji is defined as the i th input to node j, and wherein φ'(y j ) and ∂Γ/∂x ji are calculated according to the equation: (1) for a SUM node φ'(y.sub.j)=1 and ∂Γ/∂x.sub.ji =1; (2) for a LINEAR THRESHOLD node ##EQU11## (3) for a PRODUCT node φ'(y.sub.j)=1 and ∂Γ/∂x.sub.ji = .sub.k=1→n x.sub.jk where k≠i; (4) for a WEIGHTED-SUM-GRAVITY node φ'(y.sub.j)=1 and ∂Γ/∂x.sub.ji =(Σ.sub.k=1→n x.sub.jk /w.sub.jk).sup.-1 -(1/w.sub.ji)×(Σ.sub.k=1→n x.sub.jk)×(Σ.sub.k=1→n x.sub.jk /w.sub.jk).sup.-2 wherein z i is the output of node i, w ji is the weighted connection between node j and node i, and x ji is the product (z i ×w ji ).

Claim 4

Dependent Claim 172 words

The method of claim 2 wherein the error signal for an output node j, e j , is generated in accordance with the equation: e.sub.j =I.sub.j -z.sub.j where I j represents the ideal or desired output from output node j, z j represents the actual output generated at output node j, and the error signal at a non-output node j is generated in accordance with the equation: e.sub.j =Σ.sub.k=1→n ε.sub.kj ×e.sub.k, where e k represents the error value at each node k that receives input from the non-output node j, and ε kj is an error distribution weight that is calculated according to: ε.sub.kj =φ'(y.sub.k)[∂Γ(x.sub.k1, . . . , x.sub.kn)/∂x.sub.kj ]w.sub.kj, where φ'(y k ) is the derivative of a node k's output function, ∂Γ(x k1 , . . . , x kn )/∂x kj is the partial derivative of a node k's combining function with respect to the input value node k receives from node j, and w kj represents the weight of the connection between non-output nodes j and k.

References Cited (Backward Citations)

US4326259A Cited

US4593367A Cited

EP0295876A2 Cited

US4979126A Cited

US5255344A Cited

US5168549A Cited

US4912647A Cited

US4912654A Cited

US4912649A Cited

US4912651A Cited

US4914603A Cited

US4912652A Cited

US4912655A Cited

US4912653A Cited

JPH02292602A Cited

US5175797A Cited

JPH0492901A Cited

USPTO Annuity & Maintenance Windows

Annuity Stage	Window Opens	Payment Deadline	Surcharge Period Until	Estimated Fee	Status
3.5 Years (Year 4)	1999-06-04	1999-12-04	2000-06-04	$2000.00	Window Closed / Lapsed
7.5 Years (Year 8)	2003-06-04	2003-12-04	2004-06-04	$3760.00	Window Closed / Lapsed
11.5 Years (Year 12)	2007-06-04	2007-12-04	2008-06-04	$7700.00	Window Closed / Lapsed

Assigned Categories & Tags

No categories assigned to this patent yet.

Assign Category Tag

Select Category

Notes (Optional)