The digression on variability and its genetic control sets the stage to consider the issue of evolvability in a biological context. If the expression of genetic variation is itself under genetic control, is it conceivable that species evolve "strategies" of how to structure the phenotypic effects of mutations? Or, to be more precise, is it possible that evolvability is systematically produced by the evolutionary dynamics of genetic variation for variability? And does evolution produce trends in the variational properties of the genotype-phenotype map? What exactly is evolvability and what influences its degree?
Evolvability is the genome's ability to produce adaptive variants when acted upon by the genetic system. This is not to say that the variants need to be "directed" (a la Foster and Cairns, 1992) for there to be evolvability, but rather, that they cannot be entirely "misdirected", that there must be some small chance of a variant being adaptive. The situation is analogous to obtaining a verse of Shakespeare from monkeys banging away on typewriters. Typewriters make this far more likely than if the monkeys had pencil and paper. The typewriters at least constrain them to produce strings of letters. Similarly, the genotype-phenotype map constrains the directions of phenotypic change resulting from genetic variation.
Evolvability has its counterparts in various fields of computer science such as heuristic search, genetic algorithms, and genetic programming. In each of these fields the same problem occurs: one is searching a large set of objects (such as genotypes, programs, or combinations of parameters) for the objects that best fulfill some measure of quality (such as fitness, performance, efficiency, etc.), and one wishes to use the samples taken so far as a guide to what samples to take next, so that one is not merely doing random or exhaustive search. Usually the set of possibilities is too large to be searched exhaustively. As a consequence success depends on some kind of heuristic hint, an Ariadne thread, which guides the researcher, the algorithm, or the population through the maze of possibilities.
The Darwinian heuristic is to choose sample points by perturbing the more fit ones among those sampled thus far. Implicit in the Darwinian heuristic is the notion of perturbation, and the assumption that the fitness function is not completely randomized by a perturbation (thus the genome is not a "House of Cards" (Kingman 1978) in which any genetic alteration brings it tumbling down).
The paradigmatic image for successful Darwinian search is Wright's image of the population walking up the side of a "fitness peak" (Wright, 1964). If one wants to find the highest point in a landscape and one can not see far into the distance, the best guess is to walk uphill. This will lead to at least one of the high points in the landscape, but of course not necessarily to the highest point. A population is slowly accumulating better and better mutations in a stepwise fashion. However, whether this approach is successful depends on whether the shape of the fitness function with respect to the genetic perturbations actually provides the information necessary to find the best genotype or the best solution to a technical problem.
Within computer science a growing body of theory has been developed which tries to pin down exactly why certain search problems are difficult and others are easy for the Darwinian heuristic. The concepts include the ideas of deceptiveness (Goldberg, 1987) and ruggedness (Kaufmann, 1989) of fitness landscapes, epistasis variance (Davidor, 1991), and the idea of strong causality (Rechenberg, 1994) to name a few. Here we want to mention but three of these concepts which all point to the same direction.
The idea of strong causality comes from physics but is used extensively in evolutionary strategy (ES) research to explain ES performance (Rechenberg, 1994). "Strong causality" simply means that small changes in the system parameters shall, on the average, correspond to small changes in system performance (fitness). If this is the case it is easy to find a path towards the best or at least a good solution of a problem.
Similar ideas have been developed in genetic algorithms theory. The classical idea of heritability appears in correlation statistics used to characterize the ruggedness of adaptive landscapes and how far adaptation may proceed before it gets caught in a local peak (Weinberg, 1991; Stadler, 1992; MŸhlenbein and Schlierkamp-Voosen, 1995). Another approach (Jones and Forrest, 1995) measures the correlation between fitness or performance and the distance from the optimum in the search space as a predictor of how well adaptation proceeds. Evolvability is dealt with directly by generalizing Price's (1970) covariance theorem of natural selection to predict the rate at which new, fitter adaptations will be produced (Altenberg, 1995a). This rate depends on the rate of production of genetic variation by what ever means, and the correlation between the fitness of genotypes and their likelihood of producing still fitter offspring.
All these approaches are different formal ways of capturing the same intuitive notion of a (statistically) "smooth" fitness landscape: it is easy to evolve by natural selection if better genotypes are found in the mutational "neighborhood" of the good genotypes. Another way of expressing this result is that adaptations are possible if improvement can be achieved in a cumulative or stepwise fashion.
But what are the structural features that make stepwise improvement possible? The key feature is that, on average, further improvements in one part of the system must not compromise past achievements. This is the essence of the so-called "building block hypothesis" to explain the performance of genetic algorithms (Holland, 1992; Forrest and Mitchell, 1993). Independent functions shall be coded independently so that the improvement of each function can be realized with minimal interference with other already optimized functions. Pleiotropy cannot be wholly "universal" (Wright, 1968), but must be limited for many mutations. A primary problem for complex adaptation is how to avoid unbounded pleiotropy in the face of the combinatorial explosion in the number of possible interactions between parts. This is accomplished by modularity, which underlies many of the explanations of complex adaptations offered by biologists.