It strikes me that evolution by natural selection has most of the characteristics you attribute to a control system, not a selection system: feedback is far from perfect, each step of evaluation is heavily constrained by previous outputs and there is no going back, most of the search space is unreachable, it operates on the territory and there is no map, there is no final output distinct from the computation itself, and as you mentioned, it is strictly “on-line”. It’s true that it is massively parallel, and in this sense different elements of the search space are evaluated and either accepted or rejected at each “step”. I’m not sure that evolution is “obviously a selection process according to the distinction as [you] make it”.
Of course, it is an astoundingly inefficient optimizer, of whichever type it is, so it is not surprising that it lacks many of the stereotypical characteristics of its class.
It might be based on the fact that it produces agents.
I wasn’t clear on whether these was more a control thing or a selection thing—when looking at an agent, we care about what it does on its own. But we’re also interested in “evolution’s future outputs”.
It strikes me that evolution by natural selection has most of the characteristics you attribute to a control system, not a selection system: feedback is far from perfect, each step of evaluation is heavily constrained by previous outputs and there is no going back, most of the search space is unreachable, it operates on the territory and there is no map, there is no final output distinct from the computation itself, and as you mentioned, it is strictly “on-line”. It’s true that it is massively parallel, and in this sense different elements of the search space are evaluated and either accepted or rejected at each “step”. I’m not sure that evolution is “obviously a selection process according to the distinction as [you] make it”.
Of course, it is an astoundingly inefficient optimizer, of whichever type it is, so it is not surprising that it lacks many of the stereotypical characteristics of its class.
It might be based on the fact that it produces agents.
I wasn’t clear on whether these was more a control thing or a selection thing—when looking at an agent, we care about what it does on its own. But we’re also interested in “evolution’s future outputs”.