Preserving the Diversity of a Genetically Evolving Population of Nets Using the Functional Behavior of Neurons
Nachum Shamir
Current address: Department of Electrical Engineering, Technion--Israel Institute of Technology, Technion City, Haifa 32000, Israel.
David Saad
Current address: Department of Physics, University of Edinburgh, J. C. Maxwell Building, Mayfield Road, Edinburgh EH9 3JZ, UK.
Emanuel Marom
Faculty of Engineering, Tel Aviv University,
Ramat Aviv 69978, Israel
Abstract
Population diversity loss is a major obstacle in applying genetic algorithms to optimization problems, which often results in population degeneration and premature convergence. The diversity changes caused by three natural-selection strategies---comparing new offspring to the least-fit specimen in the population, to one of the parents, and to the most similar specimen in the population---are analyzed theoretically and demonstrated experimentally. Using Hamming distances, the changes in diversity induced by these strategies are analyzed for an evolving population of binary strings. The results of the analysis show that the strategy of comparing new offspring to the most similar specimen (selecting the fitter of the two) causes the smallest diversity loss.
To demonstrate the efficiency of the various methods we examine the population diversity of neural nets trained to perform certain tasks using a genetic algorithm. The functional behavior of neurons, represented by the internal representations of each neuron for the entire training set, is used to derive the functional similarity of every pair of neurons and to evaluate the similarity of every pair of nets in a population of neural networks. Using a measure of the functional behavior of neurons, the changes in diversity are demonstrated for evolving populations of nets trained on the Parity data set. The experimental results demonstrate the success of the third strategy in preserving population diversity throughout generations: overcoming obstacles in the course of training and preventing population degeneration, and thus providing more successful and reliable learning.
