[Computer-go] Tactical misevaluations in Fuego MCTS

[David Fotland]

At move 154 Many Faces is 49%, playing at R11.  Many Faces gives the 8
stones in the upper right 0.5 points for white, for a four point bias.  Many
Faces thinks the lower left is unsettled, with 0 points per stone.  

These values are based on ownership after playouts from this position,
without any UCT tree.  When you calculate ownership as below, is at based on
endpoints of playouts during a normal UCT search, or after playouts from
this position without a UCT search?

David

> -----Original Message-----
> From: computer-go-bounces at dvandva.org [mailto:computer-go-
> bounces at dvandva.org] On Behalf Of Martin Mueller
> Sent: Thursday, December 02, 2010 9:43 PM
> To: computer-go at dvandva.org
> Subject: [Computer-go] Tactical misevaluations in Fuego MCTS
> 
> At the exhibition game against Tei Meiko 9 Dan after the UEC cup, I
> resigned for Fuego when the position was hopeless. However, Fuego's
> evaluation at that time was still positive.
> 
> http://jsb.cs.uec.ac.jp/~igo/eng/result_ex/Fuego-MeikouTei.sgf
> 
> For example, at move 154 its value was 0.75 even though White clearly has
> more territory.
> The main reason for the misevaluations is that Fuego cannot reliably
> resolve some simple tactical situations. For example, the top right corner
> is a trivial semeai won by White. However the Fuego territory score for
> these points is only about -0.1, which is a bias of 0.9 per point. With 8
> stones at stake the evaluation is already 7.2 points off just from this
> region.
> 
> Another problem is the bottom left corner. This group is clearly alive but
> the evaluation is about -0.4. With about 16 points affected, the bias is
> 16* 0.6 = 9.6. Combined with some other smaller mistakes, Fuego
> underestimates its opponent's territory by 20 points.
> 
> Similar issues happened in the 13x13 games in Barcelona, and I wrote about
> it in my technical report
> https://www.cs.ualberta.ca/research/theses-publications/technical-
> reports/2010/TR10-08
> 
> How do other programs approach this issue? Is the problem as bad as in
> Fuego?
> 
> 	Martin
> 
> 
> _______________________________________________
> Computer-go mailing list
> Computer-go at dvandva.org
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go