Genetic algorithm synthesis of neural networks

• US 5,140,530 A
• Filed: 03/28/1989
• Issued: 08/18/1992
• Est. Priority Date: 03/28/1989
• Status: Expired due to Term

• Current Assignee: Honeywell Incorporated (Honeywell International Inc.)
• Original Assignee: Honeywell Incorporated (Honeywell International Inc.)
• Inventors: Harp, Steven A., Guha, Aloke, Samad, Tariq
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/329,623
• Time in Patent Office: 1,239 Days
• Field of Search: 364/513
• US Class Current: 706/13
• CPC Class Codes: G06N 3/086 using evolutionary algorithms, e.g. genetic algorithms or genetic programming
• CPC Class Codes: G06N 3/126 Evolutionary algorithms, e.g. genetic algorithms or genetic programming

Abstract

The disclosure relates to the use of genetic learning techniques to evolve neural network architectures for specific applications in which a general representation of neural network architecture is linked with a genetic learning strategy to create a very flexible environment for the construction of custom neural networks.

Claims

1. A method for synthesizing designs for neural networks which involves the use of a selected learning algorithm and a particular subject to be learned, comprising the steps of:

A. devising a bit string parametric representation of a neural network architecture having relevant parameters,

B. generating a first generation of network blueprints based on said representation which jointly include a range of values for each of said parameters,

C. generating respective neural network architectures based on the current generation of said blueprints,

D. training each of said network architectures presently defined in step C via said selected learning algorithm and said subject matter,

E. testing each of said network architectures presently defined in step C with test patterns corresponding to said subject matter for testing the receptiveness of each of said architectures presently defined in step C to the effect of said training,

F. performing an evaluation for each of said network architectures called network models presently defined in step C after said testing thereof relative to performance and cost factors of interest and assigning a score thereto representing the results of said evaluation, and wherein said network model is deemed a test model and wherein a fitness is a combined measure of its worth which takes into account at least one attribute including learning speed, accuracy and cost factors involving the size and the complexity of said tested model,

G. selecting candidates from said network architectures presently identified in step C in accordance with some rationale and applying at least one operator thereto to produce a new generation of network blueprints which shall be identified as the current generation of network blueprints based on said representation, and

H. returning to step C and continuing the process.

2. A method according to claim 1 wherein said operator is a genetic operator.

3. A method according to claim 2 further including the step of representing said architecture as a genetic blueprint via said bit string.

4. A method according to claim 3, including the further step of subjecting said blueprint to recombination relative to said range of values resulting in different but related network architectures.

5. A method according to claim 4 including the further step of subjecting said blueprint to mutation relative to said range of values.

6. A method according to claim 4, including the further step of subjecting said blueprint to mutation relative to said range of values.

7. A method according to claim 2 wherein said learning algorithm is of the backpropagation type.

8. A method according to claim 1 wherein said parameters of step A include the size of said neural network architecture in terms of computational units and the configuration of said units.

9. A method according to claim 1 wherein said neural network architecture includes computational units forming at least two layer areas, and said parameters of step A including at least one group of connections deemed a projection extending from ones of said computational units in one of said layer areas to other of said computational units in another of said layer areas.

10. A method for synthesizing, using a selected learning algorithm and a particular subject to be learned, designs for neural networks each of which comprise,

a plurality of computational units, a plurality of hierarchically arranged layer areas including input and output layer areas and zero or more hidden layer areas therebetween, each of said layer areas being defined by a number of said units, connecting means connecting source groups of said units in said layer areas other than said output layer area with object groups of said units in said layer areas other than said input layer area, said connecting means being grouped in sets deemed projections with each of said projections extending from one of said layer areas to another of said layer areas,

said method comprising the steps of:

A. providing a substring format for specifying each of said layer areas with said format having one first type part deemed a layer area parameter specifier and at least one second type part for each of said projections deemed a projection specifier,

said first type part comprising a layer area identifying address section, a total size section denoting the corresponding number of said units thereof, and a dimension section denoting the configuration formed by said units,

each said second type part being dedicated to one of said projections deemed a subject projection, said second part type comprising a target address section for identifying one of said layer areas deemed a target layer area to which said subject projection is directed, a mode of address section for said subject projection, a dimension section for denoting the configuration of an object field for said subject projection in said target layer area, a connection density section for denoting the connectivity of said subject projection to said object field, and at least one learning rule parameter section,

B. devising a bit string parametric representation of a neural network architecture based on said substring format and having relevant parameters,

C. generating a first generation of network blueprints based on said representation which jointly include a range of values for each of said parameters,

D. generating respective neural network architectures based on the current generation of said blueprints,

E. training each of said network architectures presently defined in step D via said selected learning algorithm and said subject matter,

F. testing each of said network architectures presently defined in step D with test patterns corresponding to said subject matter for testing the receptiveness of each of said network architectures presently defined in step D to the affect of said training,

G. performing an evaluation for each of said network architectures presently defined in step D after said testing thereof relative to performance and cost factors of interest and assigning a score thereto representing the results of said evaluation,

H. selecting candidates from said network architectures presently identified in step D in accordance with selected rationale and applying at least one genetic operator thereto to produce a new generation of network blueprints which shall be identified as the current generation of network blueprints based on said representation, and

I. returning to step D and continuing the process.

11. A method for synthesizing designs for neural networks which involves the use of a selected learning algorithm and a particular subject to be learned, comprising the steps of:

A. devising a bit string parametric representation of a neural network architecture having relevant parameters,

B. generating a first generation of network blueprints based on said representation which jointly include a range of values for each of said parameters,

C. generating respective neural network architectures based on the current generation of said blueprints,

D. training each of said network architectures presently defined in step C via said selected learning algorithm and said subject matter,

E. testing each of said network architectures presently defined in step C with test patterns corresponding to said subject matter for testing the receptiveness of each of said architectures presently defined in step C to the effect of said training,

F. performing an evaluation for each of said network architectures called network models presently defined in step C after said testing thereof relative to performance and cost factors of interest and assigning a score thereto representing the results of said evaluation,

G. selecting candidates from said network architectures presently identified in step C in accordance with some rationale and applying at least one operator thereto to produce a new generation of network blueprints which shall be identified as the current generation of network blueprints based on said representation, and

H. returning to step C and continuing the process, and wherein said neural network architecture includes computational units forming at least two layer areas, and said parameters of step A including at least one group of connections deemed a projection extending from ones of said computational units in one of said layer areas to other of said computational units in another of said layer areas, and

wherein there are at least three of said layer areas and at least two of said projections extend from one of said layer areas to two different ones of said layer areas.

12. A method according to claim 11 wherein said parameters include a learning rule parameter for adjusting the learning rule.

13. A method according to claim 11 wherein said parametric representation of step A is a bit string which includes the size of said neural network architecture in terms of computational units and the configuration of said units.