Fuzzy rule generation apparatus for neuro circuit network system and method thereof

• US 5,822,495 A
• Filed: 03/19/1996
• Issued: 10/13/1998
• Est. Priority Date: 04/17/1995
• Status: Expired due to Fees

• Current Assignee: LG Electronics, Inc. (LG Corporation)
• Original Assignee: LG Electronics, Inc. (LG Corporation)
• Inventors: Cho, Kwang-Bo, Wang, Bo-Hyeun
• Primary Examiner(s): Hafiz, Tariq R.
• Application Number:
  US08/618,374
• Time in Patent Office: 938 Days
• Field of Search: 395/3, 395/10-11, 395/50-51, 395/20-25, 395/27, 395/61
• US Class Current: 706/2
• CPC Class Codes: G06N 3/065 Analogue means
• CPC Class Codes: G06N 7/023 Learning or tuning the parameters of a fuzzy system
• CPC Class Codes: Y10S 706/90 Fuzzy logic

Abstract

An improved fuzzy rule generation apparatus for a neuro circuit network system and a method thereof capable of computing a fuzzy output and minimizing the number of fuzzy rules by automatically adjusting the number of fuzzy rules in accordance with a given problem and using general fuzzy function having a triangle shape or a trapezoid shape using a multi-parameter of a hidden layer and an output layer when computing a fuzzy output, which includes an input layer for receiving a plurality of data; a hidden layer for judging whether a plurality of data inputted thereto through the input layer is within an effective radius of a fuzzy variable and for automatically generating a fuzzy rule to an "n" number using the Gausian function in accordance with a result of the judgement; a parameter block for storing a parameter learned by an error back propagation learning so as to compute a fuzzy rule generated by the hidden layer; and an output layer for computing the last defuzzy operation using a fuzzy rule value generated by the hidden layer and a fuzzy variable value of a parameter block necessary for a defuzzy operation.

Claims

1. A fuzzy rule generation apparatus for a neuro circuit network system, comprising:

an input layer for receiving a plurality of data;

a hidden layer for judging whether said plurality of data inputted thereto through said input layer is within an effective radius of a fuzzy variable and for automatically generating a number of fuzzy rules using the Gausian function until each of said data is within said effective radius of one of said fuzzy rules;

a parameter block for storing a parameter learned by error back propagation so as to compute a fuzzy rule generated by said hidden layer; and

an output layer for computing the last defuzzy operation using a fuzzy rule value generated by the hidden layer and a fuzzy variable value of a parameter block necessary for a defuzzy operation.

2. The apparatus of claim 1, wherein said hidden layer and said output layer are connected by at least more than two parameter blocks.

3. A fuzzy rule generation method for a neuro circuit network system, comprising the steps of:

a first step which generates a first hidden neuron by receiving data through an input layer, sets one of said data as a center value, and sets a reference deviation value and an initial value of a parameter of a fuzzy variable in accordance with a fuzzy variable to be used;

a second step which judges whether the next data received through the input layer corresponds to a first hidden neuron, generates a new hidden neuron when the next data does not correspond to the first hidden neuron, and sets another parameter;

a third step which repeatedly performs said second step until each of said data corresponds to one of said hidden neurons;

a fourth step which learns a parameter through error back-propagation and judges whether a learned datum corresponds to one of the hidden neurons;

a fifth step which generates a new neuron when the learned datum does not correspond to said one of the hidden neurons, and sets a parameter; and

a sixth step which judges whether the number of learning iterations or errors reaches a predetermined level when the learned datum corresponds to said one of the hidden neurons, and if not, repeats said fourth step and if so, generates a defuzzy operation value by computing the hidden neuron and the predetermined parameter value which were generated in the second step and the fourth step.

4. The method of claim 3, wherein said hidden neuron Ri(X) is obtained by the following expression:

##EQU8## where Xj denoted an input data, Cj denotes a center value, and δ

² j denotes a reference deviation value.

5. The method of claim 3, wherein said second step and said fourth step is directed to judging whether an input data Xj corresponds to the first hidden neuron when an expression |Xj-Cj|² ≦

r² i) is satisfied.

6. The method of claim 3, wherein said error back-propagation learning (Ep) is achieved by the following expression:

space="preserve" listing-type="equation">Ep=1/2▪

(t-0).sup.2

where t denotes a target, and 0 denotes an output value.

7. The method of claim 3, wherein said second step is directed to generating a new hidden neuron using a hierarchically structural self-organizing learning method.

8. The method of claim 7, wherein when a new hidden neuron is generated said hierarchically structural self-organizing learning method is performed by the following steps of:

one step which sets effective parameter radii and judges whether a parameter newly generated from an existing parameter among each parameter group is within effective parameter radii; and

another step which sets a currently generated parameter as a previous parameter when the newly generated parameter is within the effective parameter radius and performs the hierarchical self-organizing learning when the newly generated parameter is not within the same.

9. The method of claim 8, wherein a condition that said newly generated parameter is within the effective parameter radii can be expressed as follows:

space="preserve" listing-type="equation">||p.sup.old -p.sup.new ||≦

Rep

where p^old denotes a previous parameter, p^new denotes a newly generated parameter.

10. The method of claim 8, wherein said hierarchically structural self-organizing learning includes a step which learns a parameter learned by a global error back-propagation and converts the previous same parameter into a current same parameter.