Jim, I'm trying to "get my arms around" what you are saying here. I'll make some probably "off the mark" comments in the hopes that you will clarify your statement...
On Sun, Jun 20, 2010 at 2:38 AM, Jim Bromer <[email protected]> wrote: > On Sun, Jun 20, 2010 at 2:06 AM, Steve Richfield < > [email protected]> wrote: > >> No, I haven't been smokin' any wacky tobacy. Instead, I was having a long >> talk with my son Eddie, about self-organization theory. This is >> *his*proposal: >> >> He suggested that I construct a "simple" NN that couldn't work without >> self organizing, and make dozens/hundreds of different neuron and synapse >> operational characteristics selectable ala genetic programming, put it on >> the fastest computer I could get my hands on, turn it loose trying arbitrary >> combinations of characteristics, and see what the "winning" combination >> turns out to be. >> > > That's a pretty interesting idea, but it won't work...I am joking, what I > mean is that it is not very interesting if you are only interested > in substantial success, it is much more interesting if you are interested in > finding out what happens. Genetic Programming has a flaw in that it is not > designed to recall outputs that might be used in a constructive combination. > The program could take the winning genome, try inverting each bit one-by-one, and observe the relative deterioration in performance. Then, the "important characteristics" could be listed in order of deterioration when their respective bits were inverted, with the most deteriorated ones listed first. That should give me a pretty good idea of what is important and what is not, which should be a BIG clue as to how it works. > If the algorithm was designed to do this, the candidate outputs (probably) > would have to be organized (indexed) by parts and associated with the > combinations that created them. > A mix of neurons with a winning genome varied slightly among subgroup(s) might potentially discover an important combination of characteristics needed for better operation, e.g. different operation for different emergent layers. Furthermore, since the output of a genetic algorithm is evaluated by a > precise method, > THIS seems to be the BIG challenge - evaluating crap. Like having a Nobel Laurette judging primary school science projects, only worse. Not only must "figures of merit" be evaluated and combined akin to end-branch position evaluation in a chess playing program, but there is added the sad fact that the programmer's (my) own ignorance is built into those figures of merit. This is why chess playing programs written by chess masters work better than chess playing programs written by really good programmers. I once wrote such a program for a commercial time sharing service and many customers played it. It never lost! It also never won!!! It played the most incredibly boring defensive game that everyone simply walked away from it, rather than spending the hours needed to ever so carefully beat it. I learned a lot from that program, NOT including how to play better chess. Hopefully I can avoid the same fate with this effort. My hope here is that the programmer (me) will become a "master" (understand self-organization) in the process, which is really the goal of this program, to train the programmer. the sense of "self organization" might be voided or at least made more > elusive and problematic. You'd have to redesign how genetic algorithms > evaluate their candidate outputs and before you did that you would have to > put some thought into how a programmer can design a test for > self-organization. It is a subtle question. > I agree. Do you have any thoughts about how to go about this? Steve ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
