Downing, Keith L. “Combining evolutionary algorithms and neural networks.” Science And Technology (2006).
URL1 URL2
Returning to the robot example of the earlier chapters, note that the various EA solutions had one key feature in common: they evolved strategies that did not change during the simulated lifetime of the phenotypes. In short, the phenotypes did not learn. In real-world situations, hard-wired solutions, whether evolved using EAs or hand-coded by humans, are difficult to trust due to the unpredictable nature of real-world environments. A strategy designed for sunny days may become disastrous in the rain, or one that assumes a flat surface is thrown into total confusion by a slight incline. In general, a good many situations call for AI systems that can adapt to their surroundings. Evolution is one form of adaptation, but it typically runs at a slower time scale than that of environmental change. Hence, the individuals produced by evolution must also have plasticity: they must be capable of changing their behavior to tackle unexpected conditions.