Evolution and learning in neural networks: the number and distribution of learning trials affect the rate of evolution

• US 5,245,696 A
• Filed: 11/21/1990
• Issued: 09/14/1993
• Est. Priority Date: 11/21/1990
• Status: Expired due to Fees

• Current Assignee: Ricoh Company Limited, Ricoh Corporation (Ricoh Company Limited)
• Original Assignee: Ricoh Company Limited, Ricoh Corporation (Ricoh Company Limited)
• Inventors: Keesing, Ronald C., Stork, David G.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/616,029
• Time in Patent Office: 1,028 Days
• Field of Search: 395/13, 395/23, 395/24
• US Class Current: 706/13
• CPC Class Codes: G06F 18/2111 by using evolutionary computational techniques, e.g. genetic algorithms
• CPC Class Codes: G06N 3/086 using evolutionary algorithms, e.g. genetic algorithms or genetic programming
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

The present invention relates to the interrelationships between nature (as mediated by evolution and genetic algorithms) and nurture (as mediated by gradient-descent supervised learning) in a population of neural networks for pattern recognition. The Baldwin effect is demonstrated that learning can change the rate of evolution of the population'"'"'s genome - a "pseudo-Lamarkian" process, in which information learned is ultimately encoded in the genome by a purely Darwinian process. Selectivity is shown for this effect: too much learning or too little learning in each generation leads to slow evolution of the genome, whereas an intermediate amount leads to most rapid evolution. For a given number of learning trials throughout a population, the most rapid evolution occurs if different individuals each receive a different number of learning trials, rather than the same number. Because all biological networks possess structure due to evolution, it is important that such interactions between learning and evolution be understood. Hybrid systems can take advantage both of gradient descents (learning) and large jumps (genetic algorithms) in very complicated energy landscapes and hence may play an increasingly important role in the design of artificial neural systems.

Claims

1. In a neural network system for use in application tasks including optical character recognition techniques, an evolutionary evaluation and learning method comprising the steps of:

forming plural genetic representations of an application task including recognizing a plurality of characters where each character has a bit configuration specifying initial connection strengths between filter lays and category units,

converting said plural genetic representations to neural networks,

assigning to each network i, k(i) learning trails, chosen randomly between zero and 2K, and

training each neural network to optimally perform the application task including optical character recognition where the number of learning trials per network is randomly chosen.

2. The system as in claim 1 wherein said training step includes gradient descent learning utilizing backpropagation.

3. The system as in claim 2 including the step of constructing a population of genes for said neural networks and converting said genes to neural networks.

4. The system as in claim 3 wherein the training of each network to perform said application task includes the step of testing each network on the task.

5. The system as in claim 4 including the step of determining whether the performance is within a predetermined percentage and if not discarding said network and if so utilizing a genetic algorithm process of crossover mutation and replication to determine new genes for the next generation.

6. The system as in claim 1 having initial parameters of N equals population size, K equals mean number of learning trials/net/generation, s equals portion of nets that survive each generation, c equals portion of surviving nets that crossover, M equals mutation rate, probability/bit/generation, and GenMax equals maximum number of generations, the method including the steps of:

initializing the population size having a size N, a mean number of learning trials per net per generation having a value K, a portion of networks that survive each generation having a value S, a portion of the surviving networks that crossover having a value C, a mutation rate, probability/bit/generation having a value M, and a maximum number of generations equal to Genmax,

randomly assigning bit values to each gene,

constructing N networks from said genes, and

presenting input/output data pairs k(i) times to said network i and performing gradient descent learning throughout said learning presentations.

7. The system as in claim 6 including the step of testing the fitness of each network, including the performance fitness on the problem to be solved.

8. The system as in claim 7 including the step of choosing the best S×

N networks to survive in order to reproduce and randomly mutating genes at rate M by choosing C×

S×

N networks to crossover mutation so as to chose the cut point randomly in the gene.

9. The system as in claim 8, including the step of generating a new population of genes Gen where Gen=Gen+1 and if Gen=Genmax, displaying the fittest network in the generation Genmax and if not, constructing N networks from said genes.

10. In a neural network system for use in application tasks, an evolutionary evaluation and learning method comprising the steps of:

forming a genetic representation of an application task having a bit configuration specifying initial connection strengths between filter layers and category units,

converting said genetic representation to neural networks,

assigning to each network i, k(i) learning trials, chosen randomly between zero and 2K, and

training each neural network to optimally perform the application task where the number of learning trials per network is randomly chosen.