137 CHAPTER 7 COEVOLUTIONARY LEARNING WITH SPATIALLY DISTRIBUTED POPULATIONS Melanie Mitchell Many approaches to machine learning require a system to learn a model from a set of training examples that the model must classify, or training problems that the model must solve. How are these training cases to be chosen? Typically, due to the costs of acqui- sition or computation, only a relatively small sample from the universe of possible problems is available to the learner. How should that sample be chosen? Clearly, if the problems are too easy or to difficult, the system will not make progress in learning. Ideally, the problems should be chosen to be optimally challenging for a system’s current state of learning and to specifically target weaknesses in the system at a given time. In life, good teachers craft such problems for their students. The challenge in machine learning is to craft such problems automatically. Approaches to this issue in the machine learning community include active learning (Cohen et al., 1996), boosting (Freund and Schapire, 1997), and coevolutionary learning. This last approach has been surprisingly successful on a number of tasks, but lacks the theoretical underpinning of the first two. It is not well understood why coevolution works, when it will succeed, and how best to apply it. The purpose of this chapter is to briefly review some of the results to date in coevolutionary learning, and to report on recent research that explores the role of spatially extended populations in the success of coevolution. Coevolutionary learning, which builds on genetic algorithms (GAs) and other evolu- tionary computation techniques, has been explored by many people. The first explicit application of computational “host-parasite coevolution” was performed by Hillis (1990). Hillis’s inspiration came from host-parasite coevolution in nature. One striking natural example is the phenomenon of “egg mimicry” in plants. Insects such as butterflies sometimes lay their eggs on plant leaves, providing a ready food source for newly hatched larvae. The passion flower plant has evolved a protection against such parasit- ism by producing toxic chemicals in its leaves. However, the genus of butterfly called Heliconius has evolved a counter-adaptation: its larvae are able to tolerate these chemi- cals. In response, the passion flower has evolved a remarkable counter-counter-adapta- tion: yellow spots on its leaves that resemble Heliconius eggs. In order to avoid too much food competition among larvae, butterflies will not lay eggs on leaves already crowded with eggs, and the spots on passion flower leaves fool at least some butterflies into thinking that there are already too many eggs on the leaves. Moreover, it turns out that the yellow spots are actually glands that produce nectar, thereby attracting ants and wasps, which also prey on the eggs and larvae (Turner, 1981). I. INTRODUCTION