Learning the Ideal Evaluation Function

Designing an adequate fitness function requires substantial knowledge of a problem and of features that indicate progress towards a solution. Coevolution takes the human out of the loop by dynamically constructing the evaluation function based on interactions between evolving individuals. A question is to what extent such automatic evaluation can be adequate. We define the notion of an ideal evaluation function. It is shown that coevolution can in principle achieve ideal evaluation. Moreover, progress towards ideal evaluation can be measured. This observation leads to an algorithm for coevolution. The algorithm makes stable progress on several challenging abstract test problems.