Sunday, August 21, 2005

 

Pondering NFL and GAFO

Pensive Pondering is a new blog and the first few posts comment on the De Jong's writing on GAFO (GAs as function optimizers) and Wolpert and Macready's writing on the no free lunch theorem. I've always been a card-carrying member of the field's optimist wing and have almost always found writing about what GAs aren't or can't do (1) trivially true and (2) uninteresting. For me the existence proof of nature's evolved and complex bounty has suggested that what genetic algorithms can do is much more interesting than what they can't. Thus, I believe we should (1) not spend one nanosecond worrying about the NFL limits and (2) not use GAFO and DeJong's thinking as an excuse NOT to explore the limits of how fast and effective GAs can be made. The book The Design of Innovation is my answer to the pessimistic wing of the field and subsequent research continues to tell us how we're just scratching the surface on our understanding of what genetic algorithms can be designed to do.

Comments:
Dr. Goldberg,

In "Genetic Algorithms are not Function Optimizers", "Part 7 Summary and Conclusions", DeJong wrote:

"
The intent of this paper is not to question the usefulness of GAs as function optimization techniques. In fact, the intent is quite the opposite. GA-based function optimizers have already demonstrated their usefulness over a wide range of difficult problems . . ."
"

Unfortunately, DeJong's paper is apparently misunderstood on this simple tenet.

I share your sentiment on usefulness, not just in GAs, but in mathematics and other disciplines.

I find many studies on the undecidable, uncomputable, etc. annoying. This is not because I do or do not believe them, but rather that I feel that humankind is in its infancy in intelligence and discovery. I'd be curious to see what the perspectives are a few hundred years from now when humankind has learned and grown.

I'm in the process of reading TDOI. This is a great summer reading book. I find that not every summer is filled with reading that is this engaging. I'm taking my time with this book as it covers a great deal and there are many simple insights that provide great thought experiments.

Thanks for sharing your positive GA experiences and what CAN be.

---O
 
Hallo!!

Very interesting the blog "Pensive Pondering".

I agree with Dr Goldberg and I think that the NFL and GAFO are much more a stimulus than a worry, because, as NFL states, there is a set of problems that some non-GA/EA can solves well, but, at same time, we need to remember that there are other problems that those same algorithms can't solve or give, at least, a good answer, so here it is a chance to use EA's. The GAFO is another interesting situation, because if we compare the results obtained by classical optimization methods (Research Operations, Monte Carlo, Hillclimbing, etc), we note that EA's, in almost all functions, show us solutions that are, at least, equal in quality when compared to solutions that those classical optimizers offer us.

And as we, in this post, are seeing GA's using a critic view, let me point an interesting article about the subject:

"How GAs do NOT Work: Understanding GAs without Schemata and Building Blocks (1995)" - Hans-Georg Beyer

That article was refused at ICGA of 1995. I think that the reason for that was the so different ideas that tried to explain GA's by another point of view and I think that the persons who analysed Beyer's article liked the things that Dr Beyer was criticizing and were not so happy with the critics, so, it is not so strange that the article had been refused(this is my opinion). The german EA school (Evolution Strategies or ES) has some critics about the three main ideas that, so far, "explain" how GA's work:

1-Schemata
2-Schema Theorem
3-Building Block Hypothesis(BBH)

The germans, and some persons from GA/EA community, say that those three ideas show theoretical aspects not found in practical experiments(see, e.g., Bäck, Michalewicz, Schwefel, Rechenberg, Beyer). They have also some critics about the use of binary codification in GA's(because if your binary string is not so much long, so, you can miss some optimal sample) and show their reasons to use real codification in GA's or to just apply an ES to the problem. There are some critics about GA's as function optimizers, but what the critics do not remember is that the SGA does not solve too much problems and they forgot the origins of GA's, at least as Dr Holland thought about GA's and its applications in adptive systems. What I find so much strange is that, as the SGA does not solve so much problems, its supporters began to make some kind of Dr Moreau-like (or Frankstein-like) experiments with SGA, that is, those supporters tried a lot of new (strange?) things with the SGA, for example: Messy-GA, Greedy-GA, Delta Coding and etc. Some names could be used very well for some Horror/Terror movies. I am not saying that we do not need to try some new stuffs in GA, but what we can't forget is the original roots of GA's, that is, we need to try some modifications, but, at same time, we need also to be very careful not to build some kind of monster that, as it is "very useful" and satisfies the researcher's expectations, could be useful for nothing more and less useful for real world problems. I think that Evolution Strategies(ES) followed a very different path in its evolution, because if you see ES's have not diverged so much from the original idea. "O.k!", someone could tell me, "but you must admit that ES's were very applied only in Germany almost until the 1990's and the GA's had a range of researchers bigger than ES, so, it is not strange that there are in GA those 'monsters' that you, Nosophorus, says". I totally agree, but we should remember also that ES were built in a very interesting base and the key-concepts(global convergence, self-adaptation) of it are very useful to optimization problems.

I am not trying to find which is better, ES or GA, I think that both are very useful(and in some situations, GA's and ES's have not so much differences). I like so much GA and I think that it has a great future in its development, but you should admit that ES are very useful too.

Anyway, GA's and ES's are great evolutionary algorithms, however I think that ES's are much more elegant. :D

[]´s

Marcelo(a.k.a Nosophorus)
 
Hey i got here searching for affiliate program making money
Your sites not too bad!

You got good rankings for them keywords affiliate program making money

I would personally like to ask you to join my site its a forum about making money online , affiliate programs etc... your a good writer , hope to see you there maybe you can learn too how to make extra cash.

Thanks
affiliate program making money
 
I discuss this topic daily myself. I also have a website that talks about ways to earn money fast related things. Go check it out if you get a chance.
 
Nice post. union bank
 
I am about to show you a way that you can generate thousands of keyword targeted links back to your web site starting today!
 
Good Posts, Keep on Blogging.
Be sure to visit

Advanced Business Marketing

Expensive Marketing Secrets Free.
 
Wow. I like your blog on 1leadlightning affiliate internet marketing program.com. Let's have some more. Stop by and visit my 1leadlightning affiliate internet marketing program.com and compare sites and blogs.
 
And A Link Back To Your Web Site Excite You?
 
We'll start delivering traffic to your website instantly.
 
Hey, you have a great blog here! I'm definitely going to bookmark you! The information and Internet marketing product evaluations that you provide are great for beginner webmasters.

I have a general business degree online site/blog. It pretty much covers general business degree online related stuff.

Come and check it out if you get time :-)
 
UoP online targets mature adults who have been in the work force for at least five years. The majority of UoP online students have been career oriented for some time. By designing their online classes so that they don’t require meetings at set times, UoP online helps busy individuals advance their educations. But don’t get the wrong idea; UoP online is not set up like other distance learning programs, Dicover the detailed information on high school online class available in kansas
 
a lot of premium products and services being offered for free. What’s best is that there is a seemingly
 
Good post. It has the right type of affiliate r information. Keep it coming
 
to all of the major RSS feed directories on the Internet.
 
affiliate betting sports is great

i found here searching for the word affiliate betting sports and your site was listed high on the word affiliate betting sports
GOOD JOB

affiliate betting sports
 
This comment has been removed by a blog administrator.
 
internet web site marketing
 
Post a Comment

<< Home

This page is powered by Blogger. Isn't yours?