Non-local adaptation of artificial predators and prey

Non-local adaptation is the acquisition of general skill at a competitive task. General skill is defined as skill against a broad spectrum of opponents rather than just those an agent encountered during its training or evolution. Biological dogma suggests that evolved creatures should be adapted only to their environment and the opponents they encounter during evolution. If this dogma applies to digital evolution, then we should not observe non-local adaptation in agents trained with evolutionary computation to perform some competitive task. A number of previous studies have found non-local adaptation in prisoner´s dilemma, in a model of competitive exclusion in plants, and in a virtual robotics task. This paper examines non-local adaptation in a virtual predator-prey system. One hundred distinct predator-prey lineages are evolved for 250,000 time steps, saving an intermediate population at time step 100,000. Predators and prey from distinct lineages are placed in competition. For the four possible comparisons in which the type of one competitor is held constant and the other is varied from less to more evolved, a statistically significant increase in ability to acquire food is seen in the more-evolved agents