System and method for dynamic learning control in genetically enhanced back-propagation neural networks

• US 5,832,466 A
• Filed: 08/12/1996
• Issued: 11/03/1998
• Est. Priority Date: 08/12/1996
• Status: Expired due to Fees

• Current Assignee: International Neural Machines, Inc. (VoiceIQ, Inc. /Ewart & Palone/)
• Original Assignee: International Neural Machines, Inc. (VoiceIQ, Inc. /Ewart & Palone/)
• Inventors: Feldgajer, Oleg
• Primary Examiner(s): Hafiz, Taric R.
• Assistant Examiner(s): Rhodes, Jason W.
• Application Number:
  US08/694,414
• Time in Patent Office: 813 Days
• Field of Search: 395/13, 395/11, 395/21, 395/23, 395/24, 395/75
• US Class Current: 706/13
• CPC Class Codes: G06N 3/086 using evolutionary algorithms, e.g. genetic algorithms or genetic programming

Abstract

In the design and implementation of neural networks, training is determined by a series of architectural and parametric decisions. A method is disclosed that, using genetic algorithms, improves the training characteristics of a neural network. The method begins with a population and iteratively modifies one or more parameters in each generation based on the network with the best training response in the previous generation.

Claims

1. A method of producing an artificial neural network wherein an output response to training is dependent upon at least a parameter value and for use in a computer, the method comprising the steps of:

a) forming a plurality of groups of individual artificial neural networks, each individual in each group having a unique parameter value, the parameter values defining a first broad range of values, the individuals within a group being related to each other by the closeness of their parameter value;

b) using the computer, applying to each individual a plurality of input stimuli and their corresponding expected output responses;

c) using the computer, testing the individuals by applying to each individual at least an input stimulus and comparing at least a corresponding output response to at least a corresponding expected output response to determine the parameter value that is a best fit, wherein the best fit is assigned to the parameter value of the individual that provides a closest output response to the expected output response;

d) using the computer, assigning new parameter values to the plurality of groups of individuals based on the parameter value that is the best fit, the new parameter values defining a second range of values; and

,

e) repeating steps (b) to (d) until an individual in a group is within a predetermined tolerance of the at least an expected output response.

2. A method of producing an artificial neural network as defined in claim 1, wherein each group of the plurality of groups has at least one parameter value that is the same as a parameter in at least one other group.

3. A method of producing an artificial neural network as defined in claim 2, wherein the second range of values assigned in step (d) is a contracted range of values including a segment of the range of values of the group of the most fit individual.

4. A method of producing an artificial neural network as defined in claim 2, wherein the second range of values assigned in step (d) is at least a substantially new range of values substantially outside of the range of values of the group of the most fit individual, thereby expanding the first broad range of values.

5. A method of producing an artificial neural network as defined in claim 4, wherein the second at least substantially new range of values includes the best fit value.

6. A method of producing an artificial neural network as defined in claim 1, wherein the second range of values is a contracted range of values, the difference between adjacent values in the second range of values being less than the difference between adjacent values in the range of values.

7. A method of producing an artificial neural network as defined in claim 1, wherein the new parameter values comprise at least the best fit parameter value.

8. A method of producing an artificial neural network wherein a parameter value is determinative of training response, the method comprising the steps of:

a) providing a population of individuals;

b) grouping individuals into groups of individuals;

c) assigning a first range of parameter values to a first group of individuals, each individual being assigned a single unique parameter value that corresponds to a particular output response;

d) assigning another range of parameter values to another group of individuals, each individual in the other group being assigned a single unique parameter value;

e) repeating step (d) for all other groups, until all individuals in all groups are assigned a unique parameter value;

f) using a computer, testing, by sampling the assigned individuals'"'"' output responses to determine if an individual'"'"'s group of assigned parameter values produced an output response that is within predetermined acceptable limits;

g) using a computer, selecting those parameter values that produced output responses that are within the predetermined acceptable limits of the desired output response;

h) using a computer, assigning new ranges of parameter values to the groups in dependence upon those values determined to be acceptable in step (g), at least one of the new ranges of parameter values including at least one of the values determined to be within acceptable limits; and

,

i) repeating steps (f) to (h) until a parameter value is found that yields an output response that is acceptably close.

9. A method of producing an artificial neural network as defined in claim 8, wherein each group'"'"'s new range assigned in step (g) is one of an expanded range, a narrower range and a contracted range of values, a contracted range of values being defined as having a higher resolution than its previous range of values.

10. In a computer system having a population of back-propagation neural networks comprising a plurality of groups of individual neural networks, an iterative process of finding at least a parameter that results in acceptable training, said process comprising iterations of a series of steps, each iteration comprising the steps of:

a) testing each neural network within the population to determine a choice at least a parameter value, wherein the choice at least a parameter value is at least a parameter value of a neural network that provides a closest response to a desired output response for a particular input string of values and wherein each group has at least one parameter value that is the same as a parameter value in at least one other group; and

,

b) assigning new parameter values to the population of neural networks in dependence upon the choice at least a parameter value, the new parameter values defining a second range of values that more closely approximates the desired output response, wherein the second range of values is at least a substantially new range of values substantially outside of the range of values of the group of the most fit individual, thereby expanding the first broad range of values.

11. A process as defined in claim 10, wherein the second at least substantially new range of values includes the best fit value.