Slippers shorties

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
Henk
Posts: 5244
Joined: Mon May 27, 2013 8:31 am

Slippers shorties

Post by Henk » Sun Aug 12, 2018 10:10 am

Great game


Sven
Posts: 3649
Joined: Thu May 15, 2008 7:57 pm
Location: Berlin, Germany
Full name: Sven Schüle
Contact:

Re: Slippers shorties

Post by Sven » Sun Aug 12, 2018 10:58 am

How many years will it take to fix that?
Sven Schüle

Henk
Posts: 5244
Joined: Mon May 27, 2013 8:31 am

Re: Slippers shorties

Post by Henk » Sun Aug 12, 2018 11:47 am

Sven wrote:
Sun Aug 12, 2018 10:58 am
How many years will it take to fix that?
Don't know. Plain alha zero method takes 11000 years for tuning on my computer. So we have to be patient.


But white did not play that bad. Looks like Black played suicide chess. (Maybe using wrong sign)


First step is to replace network with an old school evaluation and see if Monte Carlo tree search works normally.

Henk
Posts: 5244
Joined: Mon May 27, 2013 8:31 am

Re: Slippers shorties

Post by Henk » Sun Aug 12, 2018 12:53 pm

Henk wrote:
Sun Aug 12, 2018 11:47 am
Sven wrote:
Sun Aug 12, 2018 10:58 am
How many years will it take to fix that?
Don't know. Plain alha zero method takes 11000 years for tuning on my computer. So we have to be patient.


But white did not play that bad. Looks like Black played suicide chess. (Maybe using wrong sign)


First step is to replace network with an old school evaluation and see if Monte Carlo tree search works normally.
I think I already found what went wrong. Wrong sign. Don't know for how long has this bug been there. All tuning time for nothing. Guess I can learn a lot from test driven development.

Played a new game but still playing lousy. But at least black not playing suicide chess. Makes same number of mistakes as white does. Might be another bug or not enough network tuning. That's why I have to replace network with normal evaluation just to see who is wrong: the network or something else.

Sven
Posts: 3649
Joined: Thu May 15, 2008 7:57 pm
Location: Berlin, Germany
Full name: Sven Schüle
Contact:

Re: Slippers shorties

Post by Sven » Sun Aug 12, 2018 5:24 pm

Henk wrote:
Sun Aug 12, 2018 12:53 pm
Makes same number of mistakes as white does. Might be another bug or not enough network tuning.
Let's be honest: you won't make a bet that there are less than five other bugs left in it, will you? :D
Sven Schüle

flither
Posts: 12
Joined: Thu Aug 02, 2018 9:16 pm
Full name: Raf Levsky

Re: Slippers shorties

Post by flither » Sun Aug 12, 2018 6:23 pm

Best game ever, reminds me my childhood, learn Leela ;)

Henk
Posts: 5244
Joined: Mon May 27, 2013 8:31 am

Re: Slippers shorties

Post by Henk » Mon Aug 13, 2018 11:03 am

Slipper without net using monte carlo tree search with alpha beta search level 1 for evaluation.
Game looks 'normal'. So problem is bad evaluation of a position by neural network.
Training examples for neural network were generated using same alpha beta search level 1 for evaluation.
So it is not the training examples. Probably much more tuning and testing of neural network required.


Henk
Posts: 5244
Joined: Mon May 27, 2013 8:31 am

Re: Slippers shorties

Post by Henk » Sat Aug 18, 2018 4:20 pm


Post Reply