A new tournament type

[Igor Gorelikov]

Finally I have invented a new type of tournaments named "browsing
gauntlet". It can be usefull if you have a rating list and arranged
participants in accordance with it.

My idea is as follows:

The first game is played vs the medium engine.
Next opponent is selected by the game result (after each game):

- upper opponent in case of a win
- same opponent in case of a draw
- lower opponent in case of a loss

For instance, the main engine have played vs participant No 5 of 10
- it will play No 4 if it won
- it will play No 6 if it lost
- it will play No 5 if it drawn.

It is implemented in my IG WBTM 1.5.

Igor

[Dann Corbit]

I have a similar idea: A binary search tournament.

Take 16 engines, each very well known in strength, and separated by 100 Elo.

In the first match of 32 games, play the middle engine, #8.

If the unknown engine wins the match, have him play engine (8+16)/2.
If the unknown engine loses the match, have him play engine (1+8)/2.

In a total of 4 x 32 game matches, you will have a very strong picture of the Elo of the unknown engine.

Because the Elo of the known engines is fixed, and because 32 games is good enough for crude statistical estimation, it could also be possible to interpolate the next opponent (e.g if the engine loses all 32 games, don't try the middle but instead go near the bottom -- actual position to be decided by the 32 game Elo estimate).

I would call that tournament format linear interpolation tournament.

[Igor Gorelikov]

I have a similar idea: A binary search tournament.

Take 16 engines, each very well known in strength, and separated by 100 Elo.

In the first match of 32 games, play the middle engine, #8.

If the unknown engine wins the match, have him play engine (8+16)/2.
If the unknown engine loses the match, have him play engine (1+8)/2.

In a total of 4 x 32 game matches, you will have a very strong picture of the Elo of the unknown engine.

Because the Elo of the known engines is fixed, and because 32 games is good enough for crude statistical estimation, it could also be possible to interpolate the next opponent (e.g if the engine loses all 32 games, don't try the middle but instead go near the bottom -- actual position to be decided by the 32 game Elo estimate).

I would call that tournament format linear interpolation tournament.

Hi Dann,
I have similar ideas
with larger pool (300) with splitting it in two then each part in two then each part in two and so on.

32 is good for rough estimation (for instance, to rate a new engine).

Igor

[JohnBoyd]

I have been doing something like this for several years, to rate new engines. It works quite well (ie usually within +/-50 pts of a more in-depth rating) to play successive opponents that have established ratings until the new engine gets 3 consecutive losses. I start from the very bottom for a brand new engine, but if I have an idea it is >1900 I start from the middle of the list and go up.

I use a list of about 20 engines, each separated by about 50-100 pts. For a loss, the MAX rating of the new engine is set to +400 of the engine it lost to, for a tie, max/min set to +200/-200, and for a win the MIN is set to -400 of the engine it beat.

After 3 losses, I have a very good estimation of the new engine's strength by adding the min and the max and dividing by 2.

I use an Excel spreadsheet to make it easy...

cheers,

John