Improved Evolutionary Optimization of Difficult Landscapes: Control of Premature Convergence through Scheduled Sharing

Massively parallel compute rs afford to genetic algorithms t he use of very large populat ions, which allow t he algor it hms to at tack more difficult optimization prob lems than were feasible in t he past . Optimization perform an ce on difficult search spacesthose t hat are both vast and have large num bers of local optima-can be particularly crippled by a common problem of genet ic algor it hms: prematur e convergence of the population to a subopt imu m. In nature, however , com petit ion between like organisms prevents total convergence , and t his competition effect can also be int roduced to the standard genetic algorit hm by an extension called sharing. Shar ing forces org an isms within a nic he to compete. In past work, the size of the niche has been fixed , and calc ulated by hand for simple test problems. This paper int roduces an improvement to fixedniche sharing called scheduled sharing, which (1) allows for the app lication of sharing to complex problems where there is no definable best niche size, and (2) does not violate t he "black box" principle in t he calculation of niche size, instead borrowin g from simulated annealing an exponenti ally decreasing schedule. We show that scheduled sharing inhibits convergence and imp roves pe rformance for optimization pr oblem s that are difficult relative to t he size of the populati on used . *Elect ronic mai l address : mep«lvlsi . cs . caltech. edu tElectronic mail ad dress: smi th«lthink. com 444 Michael E . Palmer and Stephen J. Smith