4. Nonlinearity of µE. The performance measure is nonlinear, exhibiting "false peaks" and making it difficult to avoid concentration of trials in suboptimal regions.
5. Mutual interference of search and exploitation. Exploitation of what is known (generation of structures observed to give above-average performance) interferes with acquisition of new information (generation of new structures) and vice versa.
6. Relevant non-payoff information. The environment provides much information in addition to performance values (payoff), some of which is relevant to improved performance.
The schema concept suggests a coordinated array of robust procedures for meeting these obstacles. The procedures are all founded on the view that each structure is a "carrier" (or selected sample point) of each of the great number of schemata it instances. Because arbitrary structures are easily represented as strings (by using detectors or more sophisticated techniques such as the broadcast language) the resulting procedures apply to adaptation in all its forms. Once schemata have been defined, there is a natural means (p. 69) of comparing structures and apportioning credit by assigning to each schema the average of payoffs to its observed instances (compensating obstacle (2)). A small population of structures, when properly selected (pp. 139-40), can then store the relative performance rankings for very large numbers of schemata (compensating obstacle (1)). It is this broad data base vis-à-vis schemata (p. 87) which enables genetic plans to escape false peaks and other difficulties engendered by nonlinearities (compensating obstacle (4)). Recasting the search problem in terms of the space of schemata sidesteps dimensionality effects (obstacle (3)), at least for intrinsically parallel procedures such as genetic plans (p. 71). Under such plans the succession of structures generated from the data base (the current population) induces a highly parallel, diffusion-like spread of trials in the space of schemata (pp. 104-6). This takes place in such a fashion that there is:
(1) progressive exploitation of the best observed schemata,
(2) increasing confidence in the estimates of the expected payoff to the best observed schemata,
and
(3) testing of great numbers of new combinations of schemata (both newly generated schemata and new combinations of already tested schemata of high rank).
