Guided Evolution
Haven’t had time to read all the back material yet, but Shlok’s post reminded me of another point that I took away from the symposium on complexity I mentioned in my last post.
Often when discussing complex systems and emergence, one hears evolutionary dynamics offered as a way out of the maze. Instead of directly designing a system to prevail in a complex environment, we create an environment within which such a system will emerge.
My concern is this: Evolution is a wildly wasteful process. The vast majority of mutations fail. That’s fine if your time-scale is long and the cost of failure is negligible. This is how genetic algorithms work - the cost of failure is nil (culling binary strings has zero marginal cost) and current computing power allows thousands of generations to be simulated in minutes. If we start talking about weapons systems, however, what is the cost of failure and how long does it take to generate a new generation?
Related to this challenge is the idea that we know how to construct an environment that will select for what we want. Again using the genetic algorithm (GA) example, a major part of the art of GAs is finding a good heuristic - i.e. determining the appropriate selection criteria. Part of the process is basic trial and error. You try a heuristic, discover that it leads you someplace you hadn’t expected, go back and tweak it, and start the process again. When me move from the realm of computer algorithms to bending steal and building physical systems, would such a learning process become prohibitively expensive?
In short, if we think that evolutionary dynamics offer the silver bullet for overcoming the challenge of designing systems that will prevail in a complex environment, then what gives us confidence in believing we can manufacture and aim that bullet?
I think the key distinction is whether or not we hope to “control” the evolution centrally. Your final comment of “manufactur[ing] and aim[ing] that bullet” connotes a control mechanism that (according to Ashby’s Law of Requisite Variety) would have to exceed the complexity of all possible daughter states.
A more-effective mechanism (and one that could tolerate the wastefulness of the evolutionary process) is empowering the local phenomena to evolve as local conditions require. Far from “punctuated equilibrium”, this is more like a continuous dynamism of ceaseless change.
Comment by deichmans — April 17, 2008 @ 10:48 pm
Interesting distinction. Hmm…
It seems that such a mechanism would still require some centralized intention (if not control). Specifically, wouldn’t bounds need to be placed on the local phenomena? And, in figuring out what those bounds ought to be, would we not run into the trial and error challenge I described above?
I’m not sure I understand what you’re getting at regarding the punctuated equilibrium vs. ceaseless change point.
Thanks for coming by,
W
Comment by Wiggins — April 18, 2008 @ 12:11 am
I’m ruminating on the distinction between adaptive and reactive agents (in a bottom-up agent-based model). I could be mis-taking the post and comments here, but I think that Wiggins is talking about reactive agents that constitute the GA. Is deichmans talking about adaptive agents, ones that shortcut the waste of reactive agency in GAs? How do you do weapon systems experimentation using adaptive computational artifacts?
Comment by Moon — April 18, 2008 @ 10:33 pm