A cooperative training model for supervised learning in artificial neural networks
C.R. Chow · 1994
A new model for supervised neural network learning using genetic algorithms (GAs) is introduced. Artificial neural networks are computational models inspired by the neuro-anatomy of organisms. Our brains can perform certain computations such as speech recognition, pattern recognition, and motion control much faster and more robust than any existing supercomputer in the world. Therefore, neurocomputing is adopted as an alternative computing paradigm to conventional symbolic based computing for such computationally intensive problems. Genetic algorithms are optimization methods that apply the strategy of evolution in nature. GAs mimic the processes of natural selection and sexual reproduction to evolve solutions to a problem. The selection process allows more good genes or subsolutions to be passed on to the next generation. Furthermore, the reproduction step ensures mixing and recombination of good genes. Traditionally, supervised neural network learning is conducted using gradient descent algorithms which can be trapped into local minima of the search space. More recently, GAs were employed for supervised neural network learning because of their ability to explore a multimodal search space by manipulating building blocks, or schemata, instead of raw data points. Networks are usually encoded in binary strings, a population of chromosomes, on which GAs perform evolution operations such as genetic recombination and selection. Large networks are encoded as long chromosomes which expose many schemata to destructions inflicted by genetic crossover. The new model, referred to as the cotrain (cooperative training) model, reduces such destructions by using a distributed chromosome representation of a network and a cooperative mechanism to coordinate the evolutions in separate subpopulations. The new model is discussed and experimental results are presented.