[I'm working on a long post that uses fitness landscapes, so I thought I'd rerun some previous posts to get the conversation going.]
I grew up with a mountain in my backyard... literally. It wasn't that big (here in California we'd call it a hill) but back in the Ozarks it was a legitimate mountain and we owned about ten acres of it. Not the most usable of land but a lovely sight.
That Ozark terrain is also a great example of a fitness landscape because, depending on which side you look at, it illustrates the two serious challenges for optimization algorithms. Think about a mountainous area at least partially carved out by streams and rivers. Now remove all of the rocks, water and vegetation drop a blindfolded man somewhere in the middle, lost but equipped with a walking stick and a cell phone that can get a signal if he can get to a point with a clear line of sight to a cell tower.
With the use of his walking stick, the man has a reach of about six feet so he feels around in a circle, finds the highest point, takes two paces that direction then repeats the process (in other words, performs a gradient search). He quickly reaches a high point. That's the good news; the bad news is that he hasn't reached one of the five or six peaks that rise above the terrain. Instead, he has found the top of one of the countless hills and small mountains in the area.
Realizing the futility of repeating this process, the man remembers that an engineer friend (who was more accustomed to thinking in terms of landscape minima) suggested that if they became separated he should go to the lowest point in the area so the friend would know where to look for him. The man follows his friend's advice only to run into the opposite problem. This time his process is likely to lead to his desired destination (if he crosses the bed of a stream or a creek he's pretty much set) but it's going to be a long trip (waterways have a tendency to meander).
And there you have the two great curses of the gradient searcher, numerous small local optima and long, circuitous paths. This particular combination -- multiple maxima and a single minimum associated with indirect search paths -- is typical of fluvial geomorphology and isn't something you'd generally expect to see in other areas, but the general problems of local optima and slow convergence show up all the time.
There are, fortunately, a few things we can do that might make the situation better (not what you'd call realistic things but we aren't exactly going for verisimilitude here). We could tilt the landscape a little or slightly bend or stretch or twist it, maybe add some ridges to some patches to give it that stylish corduroy look. (in other words, we could perturb the landscape.)
Hopefully, these changes shouldn't have much effect on the size and position of the of the major optima,* but they could have a big effect on the search behavior, changing the likelihood of ending up on a particular optima and the average time to optimize. That's the reason we perturb landscapes; we're hoping for something that will give us a better optima in a reasonable time. Of course, we have no way of knowing if our bending and twisting will make things better (it could just as easily make them worse), but if we do get good results from our search of the new landscape, we should get similar results from the corresponding point on the old landscape.
In the next post in the series, I'll try to make the jump from mountain climbing to planning randomized trials.
* I showed this post to an engineer who strongly suggested I add two caveats here. First, we are working under the assumption that the major optima are large relative to the changes produced by the perturbation. Second our interest in each optima is based on its size, not whether it is global. Going back to our original example, let's say that the largest peak on our original landscape was 1,005 feet tall and the second largest was 1,000 feet even but after perturbation their heights were reversed. If we were interested in finding the global max, this would be be a big deal, but to us the difference between the two landscapes is trivial.
These assumptions will be easier to justify when start applying these concepts in the next post in the series. For now, though, just be warned that these are big assumptions that can't be made that often.
No comments:
Post a Comment