CSVN Programmers' Tournaments May 2014

[flok]

Congrats to Richard as well!

One question for Folkert about PuppetMaster, what were the seven engines running on each of the seven RPIs?

Thanks,
Gerd

Gerd,

The Udoo was running a program of myself: it coordinates the selection of moves etc and the communication with the rpi's.
The rpis ran:
# node1: bikjump
# node2: gnuchess
# node3: fruit
# node4: toga2
# node5: stockfish
# node6: minkochess
# node7: arasan

[Gerd Isenberg]

Congrats to Richard as well!

One question for Folkert about PuppetMaster, what were the seven engines running on each of the seven RPIs?

Thanks,
Gerd

Gerd,

The Udoo was running a program of myself: it coordinates the selection of moves etc and the communication with the rpi's.
The rpis ran:
# node1: bikjump
# node2: gnuchess
# node3: fruit
# node4: toga2
# node5: stockfish
# node6: minkochess
# node7: arasan

Thanks Folkert,
can you elaborate a bit how your 7-hirn master chooses the move it plays, checking scores and pvs?

[flok]

can you elaborate a bit how your 7-hirn master chooses the move it plays, checking scores and pvs?

Yes, it is rather simple:

- every node gets asked for a move in a certain position
- then it is counted how often each distinct move was chosen
- then, the most chosen move is returned
- if two moves have the same maximum count (e.g. 3, 3 and 1), then the one is chosen for which the engines have been chosen most
to clarify this: I keep track how often the move of an engine was selected, then

If all goes well in one of the next CSVN.nl magazines a small article will appear explaining all of this.

[Henk]

I used a similar idea a few years ago with predicting stock quotes. But it failed. The total did not predict much better than the units. I think it was because of correlation and perhaps because all units used the same source of information.

[flok]

I used a similar idea a few years ago with predicting stock quotes. But it failed. The total did not predict much better than the units. I think it was because of correlation and perhaps because all units used the same source of information.

I doubt that you can compare predicting stocks with playing chess.

[Henk]

I don't know. I used neural networks. Each predictor was a neural network.
Only the weights and the configuration of the units were different. Sometimes I also used other learning methods.

If you have a group of engines that play similar but worse moves than the better move of another smaller group of engines the worse move is played.

It's best that units are as independent and uncorrelated as much as possible. Perhaps using total different algorithms. But in practice that won't happen. So perhaps their only difference is their tuning weights.

Maybe it's best to compute a weighted average using ELO of the units. But I used a weighted average too.

Perhaps the more units you have the more chances of hidden "interdepencies" or correlations.

[Gerd Isenberg]

can you elaborate a bit how your 7-hirn master chooses the move it plays, checking scores and pvs?

Yes, it is rather simple:

- every node gets asked for a move in a certain position
- then it is counted how often each distinct move was chosen
- then, the most chosen move is returned
- if two moves have the same maximum count (e.g. 3, 3 and 1), then the one is chosen for which the engines have been chosen most
to clarify this: I keep track how often the move of an engine was selected, then

If all goes well in one of the next CSVN.nl magazines a small article will appear explaining all of this.

Hmm, let say Stockfish finds a sacrifice with huge winning score but all the other engines agree in having no clue but propose the same "obvious" move with much lower or even negative score, they overrule Stockfish?

[flok]

[quote="Gerd Isenberg"Hmm, let say Stockfish finds a sacrifice with huge winning score but all the other engines agree in having no clue but propose the same "obvious" move with much lower or even negative score, they overrule Stockfish?[/quote]

Yes!