Читать книгу The Uses of Diversity - David Ellerman - Страница 11
ОглавлениеIntroduction
In the late nineteenth and early twentieth centuries, evolutionary analogies smacked of the cut-throat, survival-of-the-fittest, and red-in-tooth-and-claw imagery of social Darwinism and free-market economics. Today this has much changed due in part to revolutions in biology itself and in part to the modeling of “biological” processes using very different substrata in the literature on artificial life and complex adaptive systems. We are freed to use biological metaphors in many different fields. Dynamic processes first discovered in biology might be abstractly formulated and be found to have applications in quite different fields (e.g., genetic algorithms or predator-prey dynamics). It could also work in reverse. We might first isolate fundamental processes at work in other areas and then look to see if the “blind watchmaker” (Dawkins 1986) of evolution had, perhaps, found similar mechanisms long ago.
My topic in this chapter is the process of parallel experimentation, which I take to be a process of multiple experiments running concurrently with some form of common goal, with benchmarking comparisons made between the experiments, and with the migration of discoveries between experiments wherever possible to ratchet up the performance of the group. The thesis is that this is a fundamental scheme to enhance variation, innovation, and learning in contexts of genuine uncertainty.
The use of a parallel-path strategy for the solution of difficult development problems is standard practice in several of our outstanding industrial laboratories. It is extremely common in agricultural and medical research. And in the atomic-bomb project, one of the most spectacularly successful military projects the United States has ever undertaken, the parallel-path strategy was employed. (Nelson 1961, 353)
This set of ideas about innovation and learning keeps popping up in different fields including evolutionary theory, so my aim here is to try to draw out the analogies and triangulate on the ideas.
The theme of parallelism has received renewed attention in the modern complexity sciences, for example the brain-inspired theories of neural networks and parallel distributed processing.
[The] theme under the aegis of complexity is the emphasis on parallel (network) rather than serial (hierarchical) systems. . . . The distinction between serial and parallel systems is quite general. The serial system can be generalized to a hierarchical system like the pyramidical organization chart for a corporation, the church, or the military. Hierarchical systems are such that there are a “top” and a “bottom” at every level. . . . Parallel systems generalize to what I will call a “network.” A network has no “top” or “bottom.” Rather it has a plurality of connections that increase the possible interactions between components of the network. Most real system are mixtures of hierarchies and networks. (Pagels 1988, 50)
It will be useful, however, not to cast our net too widely. As considered here, parallel experimentation assumes enough of a common goal—a cooperative aspect—so that benchmarking between experiments is meaningful and discoveries could be usefully communicated between experiments to ratchet up the whole group.1 Perhaps the borderline case is the common experimental design to concurrently test a number of already-delineated treatments on random samples of individuals to decide which is best. Ronald A. Fisher’s F-test provides a type of benchmarking between the treatment groups to see if the variance between groups is significantly different from the variance of individuals within groups.
Sewall Wright’s Shifting Balance Theory of Evolution
Sewall Wright (1889–1988) together with the same Ronald A. Fisher and J. B. S. Haldane were the three progenitors of one of the revolutions in modern biology, the mathematical theory of population genetics (see Provine 1971). In the recent complexity science literature, Wright is more often mentioned as the inventor of the “fitness landscape” to represent optimization on a very rugged and cloudy landscape. Yet the fitness landscape was only a tool Wright used to expound his shifting balance theory of evolution.2
Natural selection is a mechanism to push a population up a fitness hill—but it may be a very low hill. “The problem of evolution as I see it is that of a mechanism by which the species may continually find its way from lower to higher peaks in such a field” (Wright 1932; reprinted in Wright 1986, 163–64). How does evolution ever get the population back down a hill and across a valley of low fitness to climb a much higher hill? If selection operates to cut down variety to the survival of the fittest, what is the mechanism to increase variety in order to find a path from low to higher hills? Many biologists, Fisher and Haldane among them, don’t think any special theory is required. They have faith that the variety introduced in the whole population by mutation, sexual reproduction, genetic drift, and changes in the environment will suffice. Sewall Wright was not satisfied with that explanation.
Like Darwin, Wright thought it relevant to carefully observe artificial selection. Wright found that breeders do not keep all their animals together in one interbreeding herd. They deliberately break the herd up into subherds, subpopulations, “races,” or “demes” (as in demography). It is a question of balance. The subherds should be small enough so that the variety found in the subherd (through sampling error) or created through mutation, sexual reproduction, and genetic drift will be emphasized through inbreeding. But the subherd should not be so small that inbreeding leads to the quick fixation of ill-adapted genes and the deterioration or demise of the subherd. When a clearly superior example is produced in a subherd, then the seed is crossbred into the other subherds to give them the benefit of the innovation. But seeds could not be constantly crossbred between the subherds as that would defeat the benefits of their semi-isolation. Shifting balances were involved. How small to make the subherds and how much crossbreeding between the subherds?
Seeing these processes at work in artificial breeding and selection, Wright reasoned that Nature might have found some version of parallel “experimentation” with naturally forming subpopulations and cross-fertilization by migration.
Judging from animal breeding, (Wright) thought that natural populations must be subdivided into small-enough partially isolated subgroups to cause random drifting of genes but large-enough subgroups to keep random drifting from leading directly to fixation of genes, for this was the road to degeneration and extinction. Mass selection within subgroups was followed by selective diffusion from subgroups with successful genetic combinations. The final step was the transformation of all subgroups by the immigration of organisms with a superior genotype and subsequent crossbreeding. (Provine 1986, 236)
Using the terminology of the field, a text in population genetics describes the theory as follows:
In the shifting balance theory, a large population that is subdivided into a set of small, semi-isolated subpopulations (demes) has the best chance for the subpopulations to explore the full range of the adaptive topography and to find the highest fitness peak on a convoluted adaptive surface. If the subpopulations are sufficiently small, and the migration rate between them is sufficiently small, then the subpopulations are susceptible to random genetic drift of allele frequencies, which allows them to explore their adaptive topography more or less independently. In any subpopulation, random genetic drift can result in a temporary reduction in fitness that would be prevented by selection in a larger population, and so a subpopulation can pass through a “valley” of reduced fitness and possibly end up “climbing” a peak of fitness higher than the original. Any lucky subpopulation that reaches a higher adaptive peak on the fitness surface increases in size and sends out more migrants to nearby subpopulations, and the favorable gene combinations are gradually spread throughout the entire set of subpopulations by means of interdeme selection. (Hartl and Clark 1997, 259)
The point is that by dividing the population into demes or races, there is more variation and exploration. Since the results can be reaped by the whole population through crossbreeding, the overall rate of advance is increased.
