Tuesday, August 12, 2014

NFL&I

My last post showed that a claim of Winston Ewert, William Dembski, and Robert J. Marks II to groundbreaking research does not withstand a minute of Googling. In the comments section, I made some admissions about my own work on "no free lunch" theorems for optimization. I want to highlight them here.

If I had a face-to-face chat with Winston, I'd tell him that I really, really, really regret having gotten into the "no free lunch" thing. Rather than play to my strengths — I'm a crackerjack computer scientist — I wasted a lot of myself on obfuscated reinvention of a probabilist's wheel. What's dreadfully embarrassing is that it took me many years to realize how bad my work was. I had to acknowledge that Haggstrom [cited below] was correct in his criticism of NFL. And Winston should acknowledge that he does not have a theoretical bone is his body.
Quoting what I said about myself, Anonymous asks:
Would you be willing to elaborate on this? It might help others avoid the types of mistakes you want them to avoid, if you could explain specifically: 1) in what ways your work was poor, or a "reinvented wheel", 2) why you did not notice your work was bad, and 3) what work/areas should be avoided (for fear of reinvention). Thanks.
The following is my response, lightly edited.


The first two questions are entirely legitimate. The last borders on the fallacy of the complex question. I'm more than willing to elaborate. But I'm going to limit my response, because I'm struggling to get a couple other things written. [Obviously, I've drifted off-task.]

1) I have toiled and toiled over a preface [now available] for my first NFL paper (1996), discovering errors in my explanations of errors, and compounding my embarrassment. Dog willing, and the Creek don't rise, I will eventually post "Sampling Bias Is Not Information" here. You can arrange for Google Scholar to alert you when the title appears, if you don't want to follow my blog. For now, I'll say only that what I call "conservation of information" is nothing but statistical independence of the sampling ("search" or "optimization") process and the sample of random values of the objective function. I describe statistical independence in the introduction to the paper, but fail to identify it. That is indeed bad.

2) For one thing, I had given my confusion a fancy name. But I would rather focus here on my runaway arrogance. I independently proved the main NFL theorem in 1994. Questioning a student during his thesis defense, I got him to state clearly that "future work" was supposed to lead to a generally superior optimizer. After a centisecond of thought, I sketched the simple argument that Häggström gives in "Intelligent Design and the NFL Theorems" (2007). I considered going for a publication, but, as obvious as the result was to me, I had to believe that it was already in the optimization literature. Twenty years ago, a lit review in an unfamiliar field was a big undertaking. And I had more than enough work to do at the time.

When Wolpert and Macready disseminated "No Free Lunch Theorems for Search" through GA Digest (early 1995), stirring up a big controversy, I concluded that I was remarkably clever, rather than that the whole affair was silly. The accepted version of my first NFL paper included my simple proof, but a reviewer's suggestion led me to see exchangeability (Häggström provides the term) as the "NFL condition." Rather than save the result for another paper, I dumped it into Sect. 3.2 about the time that camera-ready copy was due. The upshot is that I obsoleted what is known as the NFL theorem, 16 months before it was published. (I sent a copy of my paper to Bill Macready, who responded, "Nice work, Tom.")

As the reputation of NFL grew, so did my arrogance. I forgot what I had suspected in the beginning, namely that something so simple was probably not a novel discovery. The more work I did with NFL, the more it became "my thing." In 2000 and 2001, I gave tutorials on the topic at conferences. I put NFL in the cover letters of my applications for academic positions. When Dembski came along with his book No Free Lunch (2002), I was able to respond with "authority." And the number of citations of Wolpert and Macready (presently 3823) continued to go up and up and up. "Fifty Million Frenchmen Can't Be Wrong," you know, about the importance of NFL.

3) Do not presume to contribute to a field that you have not bothered to study in depth.

No comments :

Post a Comment