Author: vladan
Date: 05:56:46 12/15/05
Go up one level in this thread
On December 15, 2005 at 02:35:39, Chrilly Donninger wrote: >Its obvious that the evaluation influences the style of a programm. But >interestingly the shape of the search tree has also a significant influence. As >a rule of thumb: The programm prefers lines with larger (sub-)trees. If one >increases a certain extension the chances for such moves to be played increases >(to opposite holds for pruning). >I am certain, that this effect exists, but I do not exactly know the reason. One >explanation is: The evaluation consists of a true term which properly reflects >the value of the position and white noise. If the programm has many choices, >because the search tree is large, the expected value of the white noise is >higher. For non statisticans: If one picks from a bag with 1000 numbers randomly >1 number and the next time 10 numbers, the chances are very high, that the >maximum of the 10 numbers is greater than the number picked first. > >This effect takes of course only place, if the true values of moves are similar. >In this case the move with the highest white noise is choosen. > >One example for the practical consequences can be found in Rybka: Probably a lot >of users think, when they choose the personality "very positionally", that the >programm has - in contrast to "very tactically" - more chess knowledge, that >there is a trade-off between knowledge and search-speed. >In fact the 4 personality settings change 2 numbers. These numbers influence >only the pruning/extension mechanism of the search tree. The "very tactical" >settings prune less than the "very positional" one. I have not tested the >differences in the playing styles, but from the theoretical considerations it >should indeed influence the style. The terms "tactical, positional" are just >labels. One has to give it a name. >The same is probably done in all other programms. E.g. an old Nimzo-version of >mine had already such a setting. The personalities were called aggressive, >solid.. These settings changed the shape of the tree in a similar way. Maybe >Rybka is a Nimzo clone :-) > >Chrilly Probably, it is the most important task to find balance between the complexity of the evaluation function and search capabilities. Generally, the more complex and slower evaluation function - the lower depth. But the calculation in leaf nodes and qsearch are more accurate and precise. The ideal solution will be adaptive (depending on position) evaluation function; materialistic and fast in one type of position and positional in other (closed) position. Those ideas are very old, but never implemented in sufficient way. Obviously, problem is how to determine the quality of position fast and how to decide which evaluation to use. The Lazy Eval. technique is one step inn that direction, but I am not sure that it is final (and always secure!) solution. On the other hand pruning mechanism (with same evaluation) is also important. The high pruning factors increases depth (and also miscalculations errors). The lower pruning or full width, methods generate very solid positional playing, but on lower depth. In that case, the quality of evaluation in leaf nodes became dominant. Of course, those academic fact’s are not the same on implementation level. For instance, history pruning works very well with Fruit and Toga II, but not so well (or not at all) in some other engines. I have the same opinion that Rybka’s style options change pruning/search parameters, not evaluation’s. Vladan Vuckovic
This page took 0 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.