"Evolved" Caches Could Speed the Net

SpaceDilbert writes "According to New Scientist, evolutionary algorithms could make many network caches twice as efficient. This article describes a study carried out by a US researcher and two German academics, who "evolved" algorithms to determine what data should be held at a cache and for how long."

Comments

LRU Rules

[Mirk]

There's a good reason why LRU caching (least recently used) is so widespread, and that is that it's very very hard to come up with a sophisticated algorithm that outperforms this very naive one.

For the uninitiated, elements are added to an LRU cache until it fills up; thereafter, whenever a new element is added, space is made for it by throwing away the least-recently used one. Note, least recently used, not the least recently added, i.e. the oldest, since an element that was cached long ago may be used all the time, and so be well worth its place in the cache. For example, consider the company-logo image that your browser caches when you visit a new site and that is embedded in every page on that site. However old it gets, it's likely to continue to be used while you're on the site. As soon as you move to another site, it gradually shuffles its way down the deck until it falls off the bottom - which is precisely what you want.

A note on hill-climbing

[Animats]

Genetic algorithms are methods for optimization in bumpy spaces. The basic goal is "find x such that f(x) is maximized". As optimization algorithms, they should always be tested against the two simple optimization algorithms - basic hill climbing, and random search.

If the search space is not dominated by local maxima, basic hill climbing (go for the best neighboring value) will work. And it will be fast. If the function is differentiable, it can be orders of magnitude faster than other methods, because you can use a variant of Newton's Method.

If the search space is small, random search (just guessing) will work by exhaustively searching the space. This is obvious, but tends to be ignored in academic papers all too often.

This discussion also applies to neural nets and simulated annealing.

Now this article at least describes a problem for which a GA might actually be useful. Many such articles don't. But they haven't demonstrated that you need a bumpy hill-climbing algorithm.

This is why, despite all the hype, GAs, neural nets, and such aren't used all that much. The search space has to have the right properties. Not too small, not too big, bumpy, but not too bumpy.