Friday, March 09, 2007

Marvin Minsky's View Upon Genetic Algorithms

I found an old and interesting post from the Grey Thumb blog in which the famous computer scientist and artificial intelligence pioneer, Marvin Minsky, gives his opinions on the current state of artificial intelligence, the progress it have made, and some critiques about some "non-traditional" A.I. approaches, such as genetic algorithms, genetic programming, and neural nets.

Below there is an excerpt from his lecture: It's 2001, where's HAL? - In audio format: here (to play it on your audioplayer) or here (to download the mp3 file). Here you are the excerpt:

"Genetic algorithms are very popular. I can't figure out why because, in almost all respects, they are worse than the traditional artificial intelligence heuristic search. What genetic algorithms do is use the computers ten thousand times faster to make lots of things to try. Then you have a competition so that the ones that succeed better in solving some problem, or faster, replace the ones that took longer. However, in real life I think that's the wrong thing. And evolution itself is screwy about this. The important thing is not to remember what led to success - or half of the thing you should remember is what worked. The other half is what are the 100 most common mistakes. When I was training for a Ph.D. in mathematics, everyone understood that in the mathematics world. If you hear a theorem, then you also want to know the 10 most likely ways that it won't apply. Of course a theorem is always true if the conditions are true, but if this was true for a compact set, is it also true for a locally compact set, and if not where is the counter example that shows why that kind of reasoning breaks down.

What evolution and genetic algorithms don't do -tell me if I'm wrong- is keep any record of why all those poor losers died. If it weren't for a almost religious, superstitious worship of imitating genetics which took 600 million years, well, [] 400 million years, to get to us from fish... You could say, boy, if [we] had kept some records of what went wrong and spent about the same amount of energy on learning how to avoid bugs, maybe it would have taken only 5 million years instead of 400? Who knows? But I think people who look at genetic algorithms and don't notice that they don't solve any problems that require deep thought should... I could go on all day?"

Oh! What a good hearted words!

It's an extremist statement. The most extremist part of his speech on genetic algorithms (GA) is the following:

"Genetic algorithms are very popular, and I can't figure out why, because in almost all respects they are worse than the traditional artificial intelligence heuristic search."

Surely, it is a strongly biased opinion, since Minsky represents the A.I. old school branch.

I wonder if subjects such as evolvable hardware would be possible through traditional heuristic search. By the way, taking into account the speech's title - It's 2001, where's HAL? - would not those traditional heuristic search methods have something to do with? Let alone the fact that, along the 1960s, there was an over-optimistic view on computers and what they would soon do in the next 20 or 30 years. Then, what did happen after 20/30 years? In a single word: Deception!

It's interesting (and strange) how Minsky has so many misconceptions upon genetic algorithms (and evolutionary computation too).

Thanks Heavens that Minsky (maybe) does not know what is a decomposable/separable function!


Also via Intelligent Machines.



Post a Comment

<< Home

Charles Darwin Has A Posse Check Google Page Rank