This paper demonstrates that running a genetic algorithm with injection of individuals from solutions to similar TSP problems can lead to better performance than running a genetic algorithm with random initialization. We also find that running GAs with information from the original problem or the adding one city problem can get better performance than running GAs with information from other kinds of modified problems. We believe this is because the same problem and the add1 problem contain all cities of the original problem and do not change the length of the edges and are therefore more similar to the original problems.
Although the approach seems feasible, much work remains to be done. Indexing, or defining similarity, remains an important issue. If we inject individuals that turn out to be dissimilar to the problem that is being attempted, the GA will quickly cull these individuals from the population and replace them with others. The system is robust and able to cope without a precise similarity measure. Furthermore, injecting only a small percentage of the population guards against the GA being led astray in these cases. We need to consider the effect of different selection schemes, recombination operators, and niching operators, for genetic search as well other search algorithms and associative memory models. Individuals need not be injected solely into the initial population. We can keep track of the performance of injected individuals and their progeny and use this information to design and inject individuals in intermediate generations.