It is currently Mon Mar 18, 2024 9:06 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 4 posts ] 
Author Message
Offline
 Post subject: AI Training: Board Point Proportions and No-Komi, Find Komi
Post #1 Posted: Thu Mar 10, 2022 11:16 pm 
Lives in sente

Posts: 723
Liked others: 1018
Was liked: 30
Rank: BGA 3 kyu
KGS: Elom, Windnwater
OGS: Elom, Elom0
Online playing schedule: The OGS data looks pretty so I'll pause for now before I change it.
For those who aren't familiar with probabilistic rating systems:

In chess, a win is 1 board point, a loss would give you 0 board points, and a draw would give you 0.5 a board point.
In Go too a win is 1 board point, a loss would give you 0 board points, and a draw would give you 0.5 a board point.
In fact it's pretty much the same for any two-player contest.

However, one difference between Go and chess is that we traditionally associate better endgame play with a larger winning margin, although we don't necessarily think this way for the middle game.

In any case, this means that the proportion of board points won should be dependant on the score each game. As an example, let's say a neural net trains with no-komi, which is how I think ALL neural nets should train. In one game, black wins by 7 points. If the neural net is training using area scoring, we assume that this means black has 184 points and white has 177 points. Of course this may not be the case, but this simplifying assumption is always used because what's really being done is associating a hypothetical meaning. we then divide 184 by 361 to get black's board points, and 177 by 361 for white's. If territory scoring is used, then we divide by 360 instead of 361. The same idea applies if there's komi, since komi just changes the absolute difference in score, and were simply association proportions with score differences.

Of course, some may say, there are already go AI's that are trained to score more points these days. But the main reason I want to use this method specifically is because we can use it to derive a precise komi.

First of all, I'll again say that I really think it's best for Neural Nets to train with no komi. komi is more like an addition to the rules, but since humans haven't proven, it seems like an unnatural addition in a Neural net. I want to compare games from the olden days with a Neural Net that trained on no komi, then use that same neural net to look at new games with any komi it might use, not the other way around. Since the AI was trained on no komi, its attitude to any komi will be neutral, unlike a net trained on 7.5 points analysing a game with 5.5 points komi. Then again, I rarely actually use AI, I mainly freeload off of other people's observations, haha :lol:.

Secondly, a neural net can do half of their training on one area scoring ruleset like Chinese and another half of their training on a territory scoring ruleset like Japanese, running two independent games with each ruleset each training run. Then, after both games are finished, the bot combines the node values made in both games to an average between them, and then starts a new training run of two simultaneous games using different rulesets, but both based on the same combined node value made in the previous run. Yes, I much prefer the continous improvement in AlphaZero than the one every 100 elo in the old AplhaGo Lee.

Thirdly using no-komi training we can then train a neural net to get strong and find komi. Just take the average board point difference in it's no-komi training games with a bias towards more recent games with stronger nets, a weighted average. So the neural net can give a running value on what it thinks komi should be while it's training!

As I've implied in a recent post, however, it feels like programming at that level is for wizards, not mere mortals like me. I barely know how to even use these NN programs on my own, with all the techno-mumbo-jumbo a more savvy person would find easy, haha. My brain tells me this. Advanced Calculus? No problem. Hello World? No chance. It will probably take me a dozen years to try to make a neural net myself. However, I plan to start working on it in 2030 and finish it by 2036, along with and editor for Go, Karuta, Shogi, esChess and Xjianggqi, and call it Globalent Go AI and editor.

Top
 Profile  
 
Offline
 Post subject: Re: AI Training: Board Point Proportions and No-Komi, Find K
Post #2 Posted: Fri Mar 11, 2022 9:40 am 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
Yep, details like this are worth thinking about, and you're not the only one who has thought of such ideas :)

KataGo trains with komi, but it selects the komi on its own, by setting it to the value it thinks is most fair. It is also further randomized up or down from there, so that the net gets feedback not only how fair its current guess of komi actually is (to update its choice of komi for future games), but also gets direct data on how fair all the komis around it are, as well as a small percent of games that are made deliberately unfair by adding a *lot* of randomization (which for example can randomly result in no komi or reverse komi or double komi or even more extreme, and all the values in between). And the rules like Chinese/Japanese are randomized too, as well as the board size.

So KataGo has pretty extensive experience for all reasonable komis for multiple rulesets for all board sizes from 7x7 to 19x19, and how fair or unfair those values of komi are in all of those combinations, given its current level of play. For the current training run, there hasn't been any human input making any value of komi more privileged than any other, it's all automatically learned.


This post by lightvector was liked by: Elom0
Top
 Profile  
 
Offline
 Post subject: Re: AI Training: Board Point Proportions and No-Komi, Find K
Post #3 Posted: Sun Mar 13, 2022 4:49 am 
Lives in sente

Posts: 723
Liked others: 1018
Was liked: 30
Rank: BGA 3 kyu
KGS: Elom, Windnwater
OGS: Elom, Elom0
Online playing schedule: The OGS data looks pretty so I'll pause for now before I change it.
lightvector wrote:
Yep, details like this are worth thinking about, and you're not the only one who has thought of such ideas :)

KataGo trains with komi, but it selects the komi on its own, by setting it to the value it thinks is most fair. It is also further randomized up or down from there, so that the net gets feedback not only how fair its current guess of komi actually is (to update its choice of komi for future games), but also gets direct data on how fair all the komis around it are, as well as a small percent of games that are made deliberately unfair by adding a *lot* of randomization (which for example can randomly result in no komi or reverse komi or double komi or even more extreme, and all the values in between). And the rules like Chinese/Japanese are randomized too, as well as the board size.

So KataGo has pretty extensive experience for all reasonable komis for multiple rulesets for all board sizes from 7x7 to 19x19, and how fair or unfair those values of komi are in all of those combinations, given its current level of play. For the current training run, there hasn't been any human input making any value of komi more privileged than any other, it's all automatically learned.


:bow: That puts KataGo on my list of few reasons I would want to be rich at all (to solve water and food crises in poorer regions of the world, to make international Go tournaments, to support programs like KataGo, etc . . . Otherwise it's more money, more problems :lol:).

I hope I'm not sounding a bit too intrusive here, but how does Katago learn to play for points, in language a programming noob can understand haha :lol:? I think Golaxy or Fineart does this too.

My next thought was about a Self-Learning Algorithm like Alpha-zero that trains in Shogi, Chess and Xjianggqi, but it always trains an equal number of games between the three, rather than use an equal amount of processing power between the three. Because if a game is more complex, it would require more processing power to get to a similar level as in a simpler game, so basing it on running the same number of games between the three makes sense.

Top
 Profile  
 
Offline
 Post subject: Re: AI Training: Board Point Proportions and No-Komi, Find K
Post #4 Posted: Wed Apr 13, 2022 4:55 am 
Lives in sente

Posts: 723
Liked others: 1018
Was liked: 30
Rank: BGA 3 kyu
KGS: Elom, Windnwater
OGS: Elom, Elom0
Online playing schedule: The OGS data looks pretty so I'll pause for now before I change it.
I've decided the net should use 5 different game types that alternate equally each run with a game each of Xjianggqi, Chess, Shogi, and no-Komi go on Lentear rules and fractional board points as described above, 19x19 then 13x13, because I want to see what happens when a neural net trains purely on 19x19 on no-komi. And because it tries to win by as many points as possible, it can give a value of komi it thinks is right in fractions more precise than just 1/2, because the estimated correct komi would recent-game weighted average difference in score between black and white, which of course can be any fraction between 6 and 7. And in Lentear rules fraction board points based on a maximum of 7 territory points are used, so the exact fraction of komi between 6 and 7 counts and would represent the true value of komi.

The only problem is existing to make it, haha

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 4 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users 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:  
cron
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group