Lawrence, William. “Solving XOR and Pole-balancing problems using a multi-population NEAT.” (2020).
URL1 URL2
This work looks at the use of multi-population utilisation for Neuroevolution in evolving controllers for the XOR and pole balancing problems. The single population is split into a number of smaller populations with a lower fitness thresholds the fittest individuals from each sub-population are then seeded into the final population with a higher fitness threshold. The results so far are inconclusive that a number of smaller populations are more efficient than a single large population. Different results were compared by using the total number of generations multiplied by the population to give an approximate metric of resources required to solve the problems using different methods. Small populations of around 5–10 individuals showed comparable results in resources required to reach desired level of fitness. There are a large number of parameters to consider when designing with Neuroevolution, multi-populations increase these factors by a considerable magnitude. It may be that further changing of parameters may yield better results. Future work using novelty search may suit multi-population applications better allowing a greater coverage of the search space.