[Computer-go] AlphaZero & Co. still far from perfect play ?
sheppardco at aol.com
Fri Dec 8 04:21:34 PST 2017
You can push this farther. If we define an “error” as a move that flips the W/L state of a Go game, then only the side that is currently winning can make an error. Let’s suppose that 6.5 komi is winning for Black. Then Black can make an error, and after he does then White can make an error, and so on. So the game ends up as a Black win if the total number of errors in the game is even.
Now build a spreadsheet modeling the W/L rate predicted for Black at various error rates. You will conclude that AZ and AGZ must still have sizeable error rates. For example, if the error rate were 1% then Black would win a lot more games than White, yet AZ and AGZ are still pretty close to 50/50.
A model like this has limits, but we can be certain that there is a lot of room for improvement.
From: Computer-go [mailto:computer-go-bounces at computer-go.org] On Behalf Of ChtiGo via Computer-go
Sent: Friday, December 8, 2017 5:12 AM
To: computer-go at computer-go.org
Cc: patrick.bardou at laposte.net
Subject: [Computer-go] AlphaZero & Co. still far from perfect play ?
In go, komi allows for a very fine balancing of the winning probability, whoever the first player to move.
However, under perfect play, for a given komi, Black has a either a win or a loss before putting any stone on the board. E.g., could be a win for 6.5 pts komi and a loss for 7.5 komi.
'Close to perfect play' (is this has any sense), one would expect observing unbalanced win/loss ratios for a player, when taking Black or taking White. This would be the equivallent of the 'drawish' character of chess, observed for top programs but already at amateur level.
At the level of play of AlphaZero and AlphagoZero (~5200 ELO FWIW), AlphaZero has Win/Loss count against AG0 3-day which seems quite independent of the color (31/19 with White, 29/21 with Black, see table 1 of AlphaZero paper). In great contrast with AZ/SF in chess tournament (25/0 with White, 3/0 with Black).
Although this statistics might be insufficient, can we elaborate a bit on this and hypothetize that:
1) Commonly used 7.5 pts komi defined from human games is so far a good value up to top programs level.
2) The level of play of AlphaZero (@5200 ELO), running of big machine, is still far from perfect play and there is there is plenty of room for improvement. Nothing in view close to a winning strategy.
3) Go (with komi) is deeper than chess (very biased statement ;-), since SOA programs on SOA machines still cannot find any 'color benefit'.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Computer-go