It is currently Sat Feb 27, 2021 9:37 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 6 posts ] 
Author Message
Offline
 Post subject: Flavoured weights
Post #1 Posted: Fri Jan 01, 2021 9:11 am 
Lives in gote

Posts: 430
Liked others: 82
Was liked: 78
Rank: OGS ddk
KGS: Ferran
IGS: Ferran
OGS: Ferran
Has anyone compiled or trained weights with the characteristics of classic players? Games trained to follow the style of the Yasui school, for example; or Go Seigen.

Take care

_________________
玄 之 玄

Top
 Profile  
 
Offline
 Post subject: Re: Flavoured weights
Post #2 Posted: Fri Jan 15, 2021 9:49 am 
Lives in gote

Posts: 430
Liked others: 82
Was liked: 78
Rank: OGS ddk
KGS: Ferran
IGS: Ferran
OGS: Ferran
Someone is already doing it with chess. I've only skimmed the abstracts, so far, but about 2000 games to train. We have as many games of some of the great players...

https://lifein19x19.com/viewtopic.php?f ... ead#unread

Take care

_________________
玄 之 玄

Top
 Profile  
 
Offline
 Post subject: Re: Flavoured weights
Post #3 Posted: Sun Jan 17, 2021 7:20 am 
Lives with ko
User avatar

Posts: 196
Liked others: 42
Was liked: 31
Rank: EGF 2k
KGS: MKyle
I always assumed that even the most frequently playing pro with the longest career would not play enough games for an ai built using current techniques to learn from. Pleased to hear that I may be wrong.

I've often wondered about making a takemiya-style centre-oriented bot by training using modified rules. Perhaps the rules could give an extra point to the owner of tengen? Or apply bonuses to a wider range of central points? Maybe the player with the most stones and or territory above the 4th line gets a bonus?
I would be interested to see how much each incentive in the rules would make the bot play differently?


This post by MikeKyle was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: Flavoured weights
Post #4 Posted: Sun Jan 17, 2021 7:50 am 
Honinbo

Posts: 10661
Liked others: 3587
Was liked: 3337
MikeKyle wrote:
I always assumed that even the most frequently playing pro with the longest career would not play enough games for an ai built using current techniques to learn from. Pleased to hear that I may be wrong.


Learning by self play introduces path dependency. I would expect that today's AI bots could learn how to predict the plays of a specific player in short order, such that the predictions would generalize so that predictions in positions that the player never met are not just random. OC, the level of play of the bot at that point would be rather weak. Starting from that point the bot could be trained by self play to reach superhuman strength. Because of path dependency and, I believe, the likelihood that in most go positions there is more than one optimal play, I imagine that certain recognizable aspects of the player's style would be preserved in the bot's play.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: Flavoured weights
Post #5 Posted: Mon Jan 18, 2021 1:54 pm 
Lives in gote

Posts: 430
Liked others: 82
Was liked: 78
Rank: OGS ddk
KGS: Ferran
IGS: Ferran
OGS: Ferran
It might be interesting to see what happens with players known to have changed styles. Would AI find a common theme?
Take care.

_________________
玄 之 玄

Top
 Profile  
 
Offline
 Post subject: Re: Flavoured weights
Post #6 Posted: Mon Jan 18, 2021 9:48 pm 
Dies in gote

Posts: 52
Liked others: 0
Was liked: 5
Rank: KGS 2 dan
KGS: hakuseki
Training an AI to play like a (pre-AI era) 9-dan pro seems like a hard problem. If the AI is a pure policy with no search, then it is likely to misread many situations that the pro would read accurately. But the more you use search to correct the reading errors, the more it will tend to filter out the human-style moves that the value model judges suboptimal.

Maybe an existing AI like KataGo could be run for a fixed number of iterations (e.g. 1000 visits) and its move evaluations could be used as an input feature for a new model trained to predict the human moves.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 6 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: lightvector and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group