Out of curiosity, how long does that test suite take to run? Cheers, Adrian
On Sun, Aug 29, 2010 at 5:58 PM, David Fotland <[email protected]>wrote: > This position identified a few playout bugs, and the fixes increased the > win > rate vs Gnugo from 90% to 92%. (19x19, 8000 playouts, Gnugo 3.7.10, level > 10, Japanese rules, 2000 game test). > > Thanks, > > David > > > -----Original Message----- > > From: [email protected] [mailto:computer-go- > > [email protected]] On Behalf Of David Fotland > > Sent: Monday, August 23, 2010 9:44 PM > > To: [email protected] > > Subject: Re: [Computer-go] Tactical position > > > > I ran Many Faces for 90 seconds and it likes D3 for X with 75% win rate, > 21 > > ply PV, 1.3M playouts. > > > > Nice position. > > > > > > David > > > > > -----Original Message----- > > > From: [email protected] [mailto:computer-go- > > > [email protected]] On Behalf Of Brian Sheppard > > > Sent: Monday, August 23, 2010 8:06 PM > > > To: [email protected] > > > Subject: [Computer-go] Tactical position > > > > > > This situation taught me a lot. It is obviously a win for O, but things > > > get complicated in playouts and UCT. > > > > > > 1 2 3 4 5 6 7 8 9 > > > A - - - - X - - - - > > > B - - - O X - O O - > > > C - - - O X - O X O > > > D - X - O X X O X - > > > E - O X O X O X - X > > > F - - O - X O X X X > > > G - - - - X O O O O > > > H - - - X X X O - - > > > J - - - - - O O - - > > > X to play; O wins > > > > > > O wins this position even with X to play, and despite having a 1-eye > > > group at bottom. > > > > > > The O group at bottom is cut off, but the X group immediately above is > > > also one-eyed, and can be captured in 3 moves: B9-D9-E8. The O group > > > at bottom can be captured with J5-J8-H9-H8, so O wins by 4 to 3. > > > > > > But your playout engine will need a lot of guidance for that advantage > > > to be realized, because > > > > > > 1) O's moves can be played in any order. > > > 2) X's moves must be played in the specific order B9-D9-E8. > > > 3) Proximity don't help; after X plays J5, no rule suggests B9. > > > 4) Proximity rules hurt; after X plays H8, patterns suggest J8! > > > > > > Moreover, X can win a semeai against the O group at *top*. For example, > > > if X starts with B9, and O captures with A9, then X's A7 leaves O only > one > > > move ahead: O must play B9-D9-E8 before X plays B6-C6-A8. > > > > > > In the semeai at top, again there are problems: > > > > > > 1) O might play A6, B6, C6, or A8 filling its own outside > liberties. > > > 2) X cannot fill its outside liberties, because it has none. > > > 3) O must play in a specific order > > > 4) X can play in any order that ends with A8. > > > 5) Local patterns do not hurt X, since B9 and D9 match no patterns. > > > 6) Local patterns can hurt O, since local replies to A7,B6,C8 lose. > > > > > > Even if your program implements self-atari-vs-atari logic that help it > to > > > win the base cases of semeai, it will lose for o in almost every trial. > > > > > > Now, the left-hand side is also tricky. You would think that O's 6 > stones > > > will easily chase and kill O's two loose stones and establish life, but > > > it isn't so simple. > > > > > > 1 2 3 4 5 6 7 8 9 > > > A - - - - X - - - - > > > B - - - O X - O O - > > > C - - - O X - O X O > > > D - X - O X X O X - > > > E - O X O X O X - X > > > F - - O - X O X X X > > > G - - - - X O O O O > > > H - - - X X X O - - > > > J - - - - - O O - - > > > X to play; O wins (Diagram repeated) > > > > > > First, X's first turn is D3, which is suggested by escape-atari logic. > > > So X finds the path of stiffest resistance on many (most) trials. > > > > > > Second, in reply to D3, O has to play F2, which is not a local play. > > > (That is, it is not in the 3x3 neighborhood of D3.) Because of this, O > > > will usually *not* find the path of stiffest resistance. > > > > > > Third, that situation is likely to continue! That is, X's local > patterns > > > are often helpful at string fighting, whereas O's are often wrong. > Often, > > > the error is basically that O needs to fill outside liberties, and > these > > > are located outside the 3x3 neighborhood of X's last move. (Pebbles > > > matches patterns around the last two turns, but that just creates a > > > larger pool of potentially bad moves to play. Since O should win the > > > fight, it is O's errors that matter; X cannot make an error.) > > > > > > Fourth, as long as X is alive and making larger strings, O is getting > > > squeezed on both sides. If X happens to cut on F4, then O is likely to > > > capture something. > > > > > > Fifth, whenever the UCT tree figures out how to handle a situation, > > > X can play a forcing move someplace to distract O. For example, if O > > > figures out that F2 is a good reply to D3, then X can start with B9. > > > Now X has to reply with A9, and then O can return to D3 and force X > > > to figure out everything again. There are more than enough forcing > > > moves to push the resolution of D3 into the playouts. > > > > > > X isn't a favorite to live on the left in Pebbles playouts, but it is > > > a lot closer to 50-50 than I would like. > > > > > > To win this game, O has to live on both sides. O doesn't have to save > > > everything, but there have to be stones on both sides. So having low > > > probabilities on *either* side makes O an underdog. > > > > > > RAVE has surprisingly little effect on this. The key moves for O are > > > situational, so they often do not appear beneficial. RAVE is hurt when > > > O loses globally even though it makes a "winning" local play. For > > > example, in one test, after X played D3, O lost faith in F2 because > > > it started out with particularly bad luck on the right and won only > > > 3 out of the first 30 games. > > > > > > One of the key parameters for this position is the number of trials > > > required to find B9 in response to O's initial play of J5, J8 or H9. > > > If O needs thousands of trials to discover B9 then the UCT search will > > > have a terrible time refuting every forcing sequence that X can play. > > > > > > Programs that count semeais better than Pebbles might solve this > > > quickly. A UCT program can deal with one problem, so if Pebbles was > > > accurate at move ordering *either* side of this battle, then it > > > might be "easy" for UCT to see that O wins. But that isn't what > > > happens, in fact. > > > > > > The bottom line: Pebbles rated X as a 90% favorite when this position > > > occurred in a game. I am working through the many, many tactical issues > > > that have to be addressed to evaluate this situation correctly. > > > > > > It is a rich position because solving it will require upgrades to > > > many aspects of the system: semeai logic, ladders, big eye, nakade, > > > hole-of-three, approach moves, RAVE, and maybe reworking patterns > > > in the UCT portion of the tree. Testing has already revealed bugs > > > and new opportunities. > > > > > > Best, > > > Brian > > > > > > _______________________________________________ > > > Computer-go mailing list > > > [email protected] > > > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > > > > _______________________________________________ > > Computer-go mailing list > > [email protected] > > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
