I just went to a talk today in my institution by Professor Robert Pennock who spoke a little about education and a little about artificial evolution of intelligence. The latter stole the limelight due to the intrinsic coolness of robots (real or simulated) with goal-directed behaviour. But among his many departmental affiliations at Michigan State is the Philosophy Department so I was interested to hear what he had to say on this score. No doubt because this was a general lecture this content was mostly geared towards understanding why the argument for design is incorrect – familiar territory for evolutionists (though I liked his updating Paley’s argument from watch implies watchmaker to neat iPod app implies SDK programmer). But my parenthetically referring to robots as simulated above was deliberate. He made what I thought was a small but elegant point about this.
Is artificial evolution simulated evolution or an instance of evolution? Pennock argued that the distinction between these descriptions depended partly on pragmatics and partly on what causal processes are of interest. Since the artificial evolution paradigm he discussed modelled the minimal parameters for evolution by natural selection he was happy to describe it as an instantiation even though the mechanisms of metabolism and reproduction were not quite nature-identical or gooey enough. He pointed out that these mechanisms were not known to Darwin and yet we would still call his theory a theory of evolution.
To my mind this is equivalent to the claim that evolution is substrate-neutral. However in so far as parameters necessary for selection are dependent upon genetic mechanisms we might dispute this. In fact the metabolic and reproductive rules in the software he described entailed these parameters (often explicitly – he affirmed that recombination, a source of variation that can break up clonal interference, was modelled in some of his project). So I agree with his implicit point that modelling can be more or less precise along different dimensions such that there is an arbitrariness about the dividing line between simulation/instantiation unless a modeller’s purpose is borne in mind. But I am not sure that specifying what is modelled, and how precisely, is enough. This is because evolution is diverse and is caused by multiple processes at the population genetic level. This is clear in my field of wet-lab experimental evolution in which large populations and strong selection pressures often inflate the perceived important of selection.
Finally he was asked a semantic question about the distinction between machine learning and artificial evolution – a particularly germane one since he was describing the evolution of intelligent systems. Essentially he described the second as an example of the first in which genetical algorithms did the learning by selection. But this calls to mind an earlier distinction I remember hearing (apologies for not attributing) regarding robots that learn about their environment through feedback. An effective robot would understand its environment, i.e., whittle down its set of models, by taking actions that resulted in the maximum discrepancy between the predictions of each model. But when it came time to act in the environment, the best actions were those that minimised the predicted discrepancy between surviving models. (Something like this might be relevant to earlier efforts on this blog to explain the functions of play). I wonder if, in particularly rugged fitness landscapes, this distinction between testing and optimal inference may be important when it comes to design assisted by genetical algorithms. The ability to allow heritable variation in mutation rate may become more important in such instances.