Author: Don Dailey
Date: 20:35:50 07/19/99
I have noticed a lot of posts lately on the subject of MTD and thought I would give my observations and experiences. First of all, I would say that MTD is simply a big win. A lot of people have reported various problems with it including myself. But all the problems are correctable and you will be rewarded with the fastest possible aspriation search on the planet. All of the problems are based on not completely understanding what is going on and not bothering to stick with it until you figure it out. Even regular alpha beta searching is full of gotcha's and a lot of people don't fully understand the proper way of doing aspiration searching. This is forgivable, though, it's complicated and very easy to overlook some of the hairy issues. It's one of those things that seems ridiculously simple once you understand it, but until then is not so simple. The lazy evaluation problem is one I ran into with Cilkchess. When I tried to use lazy evaluation I got big speedups in terms of nodes per second, but the number of nodes inflated to make it NOT a win. This was quite annoying but was caused because the value you return to the mtd driver was often "weaker" because of the "cheat margin" you used with your evaluation. The solution is not to use beta (the single goal value of an mtd probe) but to use the global lower/upper bound that the mtd driver itself keeps track of. Apply your scoring margins to THOSE values because they are "true" bounds, not speculative bounds. I learned about this from discussions with others at the world computer chess championship. It was one of those things that should have been obvious to me but wasn't. I would like to mention that I was forced to use MTD and solve these problems (also problems like bad PV's) because it was simply faster, and if the speedup was trivial I would have gladly just avoided the issue! - Don
This page took 0.02 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.