Author: Uri Blass
Date: 08:43:19 10/23/00
Go up one level in this thread
On October 23, 2000 at 11:07:55, Ratko V Tomic wrote: >> GT has larger-than-life king safety scores. That >> is all. No different search paradigm or anything else. > >There are several end-user easily tunable programs. Programmer's >have also been tuning their evaluation functions for ages. Had >it been just matter of improving strength by increasing king safety >scores, most programs would be playing in that style already. Bob Hyatt did not say that gambittiger is result of changing some numbers like in chessmaster. > >To get GT's performance, there must be some additional ingredient that >other programmer's have missed. Since Christophe was talking about plans >to unify GT with regular Tiger, it has to be some extra work done and >spread out (so it can't be trivially transplanted into the regular >Tiger) within the lower level search code which helps it discover and >judge attacking chances better than other programs. > >Or perhaps that in combination with some new more elaborate king-attack >oriented post-processing evaluator (using the above mentioned info) >which affects the evaluations and the search in the next iteration. > >This would in effect create a feedback loop between the low level search >and the knowledge based post-processor, but unlike conventional leaf >evaluators, where knowledge is applied at the leaves and flows up the >tree toward the root, here it would flow both ways, where the >post-processor for the iteration N would act as a pre-processor for >the iteration N+1. That would be a much more dynamic and sensitive >pre-processor compared to the static pre-processors (which only looks >at the features of the root position). > > >In any case, however far or close the above conjectures may be, it would >be a difference of at least that size that makes GT play in the risky, >attacking style and still perform exceptionally well against other programs. > >The conventional wisdom is that such style, while effective against humans >(such as Rebel's Anti-GM mode), will fail to impress and spook other strong >programs. That conventional wisdom wasn't some arbitrary dictum. It was >the result of many years (and across many different programs) of precisely >the kind of king safety tweaks which you (and Uri) are now offering as >an explanation of the GT's unique combination of a risky style with the >superior comp-comp performance. It doesn't fit. I suggested that a different evaluation function is the reason for the fact that gambittiger sacrifies and has good results. I did not say that it is easy to find the right evaluation(part of the evaluation is knowing what to evaluate so the fact that programmers did not find a way to teach other programs to get good results with big king safety scores proves nothing). I did not claim that I know what Gambittiger evaluates in order to get big king safety scores. Uri
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.