EvoWeb Interviews John Koza
Check his interview here.
I have not been aware that John Koza was John Holland's student along the 1960s.
Interesting interview, despite my own disagreements on some points.
Koza summarizes his path towards genetic programming (GP), the choice for LISP as the main computer language of early GP implementations, the initial papers, the hostility from the GA community at that time (early 1990s), his practitioner style, his achievements, the 1000 nodes computer cluster at Stanford, and so on.
The interview is from 1998.
:.::.::.::.::.::.::.::.::.::.::.::.::.::.:
P.S: There is someone that extremely disagrees upon Koza's statements.
:.::.::.::.::.::.::.::.::.::.::.::.::.::.:
Labels: Evolutionary Algorithm, Evolutionary Computation, Genetic Algorithm, Genetic Programming, John Holland, John Koza, Old School, Simulated Evolution
posted by Marcelo at 2:50 AM
5 Comments:
Where do you disagree? Would be interesting to know...
Yes, I was student of John Holland inn 1960's. John Koza
Hi, JJ!
Sorry for the long waiting until this current reply, I was a little busy at university.
My disagreements is mainly related to Koza's statements on the theoretical approach to genetic programming and what he qualifies as "academic problems". It's important when it comes to algorithms to deal with those aspects, such as time convergence and so on. I believe that it is difficult, in some sense, to transform the evolutionary algorithms in a kind of science - or simply put them on mathematical formal grounds - if the field itself lacks the important ingredients. However, the most important, in my humble opinion, has nothing to do with simply finding a theory, a proof, a time convergence, and etc. for evolutionary algorithms, but what benefits could those algorithm features bring us. For example, what about population size reduction? Good time convergence? Self-adaptation? Less CPU time usage? Money saving? Small computer clusters? It's not simply solving a given problem, but solve it well!
By the way, thank you for your - and Koza's too - nice comments! :)
Hasta La Vista!
Marcelo
The problem with theory is to make it trickle down to practice; sometimes it's too far away (too many simplifications, for instance) to be of any use. And frankly, most of us are not mathematically prepared to understand it anyways.
Hi, JJ!
I agree with you.
Even those evolutionary algorithms which have stablished some kind of theory rely on simplifications to obtain it.
I would state that most of us even do not master the mathematical skills to draw such a theory - let alone that the inner working of some evolutionary algorithms are so complex that developing a theory for it would be humanly impossible. BUT, an initial step should be done towards that understanding.
I still believe that developing and understanding the theoretical facets of evolutionary algorithms would give us valuable benefits, despite the possibility that the cost of obtaining it may be much more expensive and useless rather than bring some advantage for the field.
While the theoretical aspects of evolutionary algorithms remain untamed, we must use what we have available, for instance computer clusters, parallelization, big populations, and so on... even though they are a little expensive.
Hasta La Vista! :)
Marcelo
Post a Comment
<< Home