Links
Chess in 2010
and if possible answer the questions you are willing to answer.Chess 2010 introduction
Introduction
This is a short recompilation of a big (and quite technical) discussion that recently took place in a public chess forum on Internet which was started by REBEL's programmer Ed Schröder.
Schröder pointed out that making a chess program to perform better in the computer-computer area doesn't necessarily mean that the program automatically will play better against humans too.
As a matter of fact Schröder believes there is even a great risk that the opposite might come true and that chess programs only become better in the computer-computer area but will lower in strength against humans.
Schröder noticed that adding more tactical power to his chess program REBEL, this in return for taking out existing chess knowledge, made REBEL a stronger computer-computer player but also that REBEL's positional understanding lowered because of the lack of the chess knowledge that was removed.
This phenomenon is not new. Hans Berliner programmer of the famous HITECH program already pointed this out some 15 years ago. He more or less proved that chess knowledge (in computer-computer play) is only worth just one extra ply.
The BERLINER experiment
Berliner took the Hitech program, removed important chess knowledge but left the very basic chess knowledge and called this version Hitech_Low (HL) and his original version Hitech_High (HH).
Then computer-computer matches were played with HL and HH at several ply-depths.
It showed up that HL (8-ply) lost from HH (8-ply) because HH was a much more intelligent program than HL, but that HL (9-ply) vs HH (8-ply) was already about equal and that HH (8-ply) was crushed when it played HL (10-ply)!
The conclusion more or less was: (extra) chess knowledge is just worth one ply in the computer-computer area. The advantage of HH having much more chess knowledge than HL wasn't enough to survive against a program with only the very basic chess knowledge searching 1-1½ ply deeper.
Recent Computer Chess developments
Especially the last years ideas based on the BERLINER experiment have been practiced in chess programs with extremely good results in the computer-computer area. Chess programmers discovered that by a new approach that looks illogical at first sight their programs became a lot stronger in the computer-computer area.
Computer chess development of the last years:Schröder about this new trend:
- Make the program as fast as possible.
- Even remove existing chess knowledge to ensure a fast chess program.
- Add lots of tactics to out-search the opponent.
- Add only the very basic chess knowledge (Berliner concept)
- Make the program aggressive to ensure tactics in games.
It's my opinion it all seem to work in the comp-comp area.
Now for comp-comp lovers this is great news but what about the people who use a chess program for analysis, study and playing games?
10 days ago I released Rebel10.0c as an engine update for Rebel10.
In 10c I removed some chess knowledge which made Rebel 30% faster, next this 30% speed gain was used to add new tactics. As a result 10c is a better comp-comp player now but a lower positional player than the original Rebel10.
This is actually what is happening the last years and I want to report it in all its details in the hope it will be understood.
There are advantages and disadvantages. The advantages are for the comp-comp lovers. The disadvantages are for people who use a chess program for analysis, study and playing games as the positional understanding of chess programs will drop and no progress is made.
My personal opinion is that the disadvantages are bigger than the advantages because the vast majority of people that use a chess program use it for playing games, analysis and study.
It's my wish the computer chess press will pick up the subject, explain and judge.
Other expert opinions
Several other colleagues of Schröder joined the discussion and gave their opinion.
Robert Hyatt, programmer of Cray Bitz and Crafty:
There is an alternative. Do as I try to do... Ignore computer vs computer games unless you see something you are consistently doing wrong in those games.
I'm not hung up on trying to beat every program by being as fast as possible. I have gotten _slower_ the past 3 years, not faster, and I think that direction is perfectly ok. I've said before that _far_ too much attention and emphasis has been placed on SSDF results, which has led to today's situation.
I think it much more interesting to continue what I've been doing for along time: trying to find ways to make these 'anti-computer' strategies backfire. Because once you learn how to 'break anti-computer' you also learn how to _play_ it. And there would be nothing to prevent you from using that against other computers, of course. :)
Bruce Moreland, programmer of Ferret:
Responding to Ed's original post
I have considered these issues carefully. My own goal is to create something that is interesting and strong, and I'm not going to gut my program's style just to get an extra quarter ply to use against computers.
I think that in the long term a good eval function is necessary to avoid catastrophic mis-evaluation of crucial features, against both humans and computers.
I avoid things like the Swedish list, in part because they've never invited me to send a version, and in part because I don't want to get sucked into that whole core-wars deal.
I would much rather be ranked on the FIDE list, which is where every player really belongs.
In Paderborn, assuming I am selected to go, I will be competing with something that will also play, with no modifications, against the humans on ICC, which is how I've always done it.
bruce
REBEL-BOARD feedback
A link-list of interesting postings on REBEL-BOARD concerning the topic.
- François Bertin
- reply to François Bertin by Ed Schröder
- Roy Brunjes
- reply to Roy Brunjes by Ed Schröder
REBEL policy
As a result of the above Rebel company feels that something need to be done for the future to ensure progress in both areas (Man vs Machine and Computer vs Computer).
To reach that goal CHESS TIGER programmer Christophe Théron from now on will only focus on computer-computer programming and make his program (if possible) the best computer fighter around.
It's our understanding that although in minority many people are in love with computer vs computer play and we want to keep serving them.
REBEL programmer Ed Schröder will concentrate to make his chess program an even better positional and human-alike chess player and make REBEL an even more attractive and configurable chess engine.
Doing so we believe we can serve the chess community best as progress is guaranteed in both areas.
For more information about the REBEL-TIGER approach CLICK HERE.
Opinion poll
Because of the importance of the subject we invite you to have a look at the CHESS 2010 POLL
Results of the CHESS 2010 opinion poll are published HERE so everybody is able to give their view.
CHESS 2010 matches
This is a new computer-computer event called CHESS 2010. The goal of CHESS 2010 is to give a preview of what one may expect from chess programs in the year 2010 concerning their playing strength in the computer vs computer area.
The idea is to play automatic games between a few top chess programs on the very high time level of one hour per move average. It's estimated that one game takes 5-7 days to finish.
Rebel company is offering 3 autoplay pairs (6 PC's) for this experiment the coming months to ensure a reasonable amount of games. The machines used for the experiment and other conditions are:
- 4 identical Pentium-II 266 Mhz with 64 Mb Ram.
- 2 identical Pentium-II 450 Mhz with 128 Mb Ram.
- Programs will play on their default (or strongest) settings.
- Programs will use the maximum hash table setting possible.
- Rebel10.0c (anti-GM=off system Tiger=off)
- Fritz 5.32
- Junior 5.0
- Shredder 3.0
- Crafty 16.6
- Hiarcs 7.0
- Nimzo 99
Database:CHES2010 *** Top 40 perc. ***
Name Tot Score + ½ - Perc
1. Rebel10.0c 9 5.0 - 4.0 2 6 1 55%
2. Hiarcs 7.0 4 2.0 - 2.0 1 2 1 50%
3. Shredder 3.0 2 1.0 - 1.0 0 2 0 50%
4. Junior 5.0 1 0.5 - 0.5 0 1 0 50%
5. Fritz 5.32 2 0.5 - 1.5 0 1 1 25%
Download games in REBEL format including analysis.
Download games in CBH format including analysis.