Early Experiments Demonstrated That Gas Often Fail 2.8 Experiment 6: Other Mutation Rates 2.4 Experiment 2: Competition between Diierent Schemata

due to convergence; not \premature convergence", or \convergence to a suboptimal string", but simple convergence. Once the population converges, either across the entire string or in a particular region, various GA mechanisms missre. The series of missteps traced in early attempts to connrm the hypotheses hint at the complications that will inevitably arise with the introduction of endogenous tness functions to our adaptive algorithms. There are exciting days ahead. On the basis of this work I ooer the following: Prescription for avoiding convergence (retaining het-erogeneity in GA populations). 1) Reward multiple schemata in the various regions of the string so that the entire population is not dominated by a single highly successful schema. 2) Use an endogenous tness function or, minimally, modify the existing exogenous tness function to share rewards among individuals with particular traits. 3) Age individuals by gradually reducing their tness. The endogenous crossover control mechanism presented here is completely general; it is blind to the particular representation in use and incurs little overhead ; it requires doubling the length of the bitstrings, counting the number of individuals with each rewarded schema and dividing the reward accordingly (these last two costs disappear with any truly endogenous tness function). Metabit crossover control is a low cost, simple means to implement diierential crossover probability and so improve performance of GAs searching for strings composed of multiple building blocks. It can be easily appended to most GAs and is likely to improve performance in cases where the populations do not converge too rapidly. Eliminating convergence through the use of dynamic endogenous tness functions and the introduction of competition may provide the pressure needed for our algorithms to evolve to successfully address diicult real-world problems. Analysis of HLA-A/B recombinant families with new poly-morphic markers. Figure 4: Crossover reducing bits. Crossover reducing values (like Figure 2) of the extreme ratio 1:20 plus a comparison of dominant and recessive metabit control. The two hypotheses were born out only for what seemed extremely high crossover rates; but it may be that they are not as high as they appear. Has an alert reader noted that 1/70 is not equivalent to single point crossover, but half that? Since a metabit and an exon bit are treated as a unit for crossover there are only 34 crossover sites in each string. Thus the crossover ratio 17:17 is equivalent to two-point crossover. As can be seen from Table …