Author: Christophe Theron
Date: 18:22:29 01/19/00
Go up one level in this thread
On January 19, 2000 at 17:23:13, Amir Ban wrote: >On January 19, 2000 at 13:01:19, Christophe Theron wrote: > >>On January 18, 2000 at 16:35:35, Dave Gomboc wrote: >> >>>On January 18, 2000 at 16:30:15, Christophe Theron wrote: >>> >>>>On January 18, 2000 at 10:32:51, Enrique Irazoqui wrote: >>>> >>>>>I noticed that sometimes Tiger keeps playing the same losing opening lines. In >>>>>the match Tiger-Hiarcs, games 6 and 10 are D10 openings, identical until 14... >>>>>e6, where Hiarcs shows an evaluation of 0.82 from its point of view. Tiger lost >>>>>them both. Games 17 and 19 were A45 lines that immediately out of book Hiarcs >>>>>evaluated as 0.57 in its favor. Again, Tiger lost both games. It seems that >>>>>there is a learner problem that will allow Tiger to play repeatedly the same >>>>>losing lines. >>>> >>>> >>>>You are right Enrique. This is not a bug, I would rather call it a design flaw. >>>> >>>>I have written a simple learning algorithm because I did not want to invest too >>>>much time in it. In fact I had almost no time to invest, so I have done my best >>>>in a very tight schedule. >>>> >>>>My philosophy has always been to put my efforts on the real stuff: the engine >>>>itself. >>>> >>>>But I'll have to work harder on the learning system, because as you have heard >>>>recently, Junior6 is badly taking advantage of this (I have heard it has >>>>repeated TWELVE TIMES the same won opening in one of the SSDF matches). :( >>>> >>>> >>>> Christophe >>> >>>Looks like you'll be doing some comp-comp testing in the future. Welcome to the >>>Core Wars saga. >>> >>>Dave >> >> >>Not necessarily. All I need is a learning algorithm that really knows how to >>avoid playing the same lost game twice, and maybe tries from time to time to >>replay a won game. That does not imply I'm going to buy 10 computers and let >>them play auto232 matches all day. >> >>Anyway I don't have enough money for that. >> >> >> Christophe > >The easiest defense against learners is to put a random element in your search >engine. If you think about it, that's not very different from what a wide book >would do for you. You are right, I should seriously think about that. It might be a quick solution to this problem. Does Junior use this? As I said, I am reluctant to spend too much time in this. I think it is mostly useful in comp-comp games, with very little benefit for the user. That's why I did not implement agressive book learning (trying to replay won openings). I think customers want us to spend more efforts in adding knowledge in our engines than in specialized anticomputers algorithms. Christophe >I think this can solve about 80% of the problem. The 100% solution is to have a >good book and a good learner. > >Amir
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.