We have found in some experiments that a useful high-level description of the case-base and the search recorded within it can be generated as follows:
One can quickly see evidence of the GA's path through the search space. Many of the earlier subtrees may have low average fitness and low order, whereas later subtrees will have higher order if convergence was approached. Typically, later nodes have higher fitness, but as clustering is sensitive to genotype, it is possible to have a weak but young subtree.
Very low-order long-lived subtrees with poor fitness are often associated with crippled individuals formed by mutation or unfortunate crossover. If there are many weak subtrees one can infer a stronger focus on exploration of new parts of the space as opposed to exploitation of strong, higher-order schema.
High-order short-lived subtrees offer evidence of temporary concentration of search effort, which often follows discovery of a beneficial building block. Once a building block has been incorporated into the population (exploited), the search effort may move on to further exploration.
In problems where there is deception (that is, where most of the gradient in the space draws the GA away from the absolute global maximum but toward some broader, more obvious local maximum) one would expect to see the GA focus first on the local maximum. If the GA manages to discover points near enough to the global maximum that they are favorably evaluated, a paradigm shift may occur. This has been evident in our case-bases.
If one notices that a small but strong subtree was discovered but quickly lost and not surpassed, one might be well advised to alter the GA parameters in favor of a more elitist selection strategy (e.g., rank selection). Note that ``Crowding'' operators have been shown to be powerful in this regard, but they add overhead and are often unnecessary. When there is suspicion of deception, it may be confirmed or discredited by looking at the lower levels of the quickly lost cluster. If one discovers that within the cluster the later and stronger subclusters had schema changes which appear to cause relative decline in other contemporary clusters then one should continue to suspect deception. At this point one may wish to more thoroughly investigate that part of the search space without much competition.