Absolute or Relative Fitness
Traditional evolutionary techniques rely on a fitness function that measures how high is a certain instance of a program in the pyramid of learning. The genetic algorithm is trying to reach a maximum fitness value.
However, in more realistic domains the calculation of an absolute fitness value must be abandoned. We want to use the computer to solve a problem for which we ourselves don't know the answer.
The technique of coevolution tries to learn by setting an arms race where partial solutions compete among themselves for partial credit. The fitness value for a particular individual is its relative success in the current generation.
As the evolutionary experiment goes on, it is difficult to have a measure of general improvement, since there is no available measure. In Tron we use only inconclusive indicators: The duration of the games (games among knowledgeable players should be longer) or the performance against a human player or a certain individual program used for comparison.