Mutation Rate Modulation for Faster GAs

Motivation and the Goal

First consider a GA implemented with the election operator, with a static search space. That is to say that the fitness function is time-invariant. As more and more generations elapse, the string population becomes more and more converged to the optimum.

When the population is fully converged on the optimum, the rate of mutation is irrelevant. Even 100% mutation probabilities will not have an effect–the election operator will not allow any new strings into the population because their fitness values do not exceed those of their parents.

If this static fitness function has the potential to become dynamic, then the a higher mutation rate may benefit the GA's "transient response" in the event that a change does occur.

As a GA reveals a fitness function, though, that fitness function is alternately static and dynamic from the string's point of view. From this it seems there should be some optimal time-series of mutation rates that decrease the expected convergence time.

Hence, the goal of this study is to determine a method for finding the optimal time-series of mutation rates for a given function, and to study its application to a handful of fitness functions.