(;GM[1]FF[4]SZ[9]AP[SmartGo:1.4]CA[UTF-8]
PW[viking5]
PB[Orego]
DT[2006-08-01]
PC[The Kiseido Go Server (KGS) at http\://kgs.kiseido.com/]
KM[7.5]
RE[W+Resign]
RU[Chinese]
C[Orego [?\]\: GTP Engine for Orego (black)\: Orego (BestPlayer1/B + BestPlayer1/W) version unspecified
]
;B[ee]CR[ee]
C[viking5 [?\]\: GTP Engine for viking5 (white)\: Viking version 5.3.1
]
;W[eg]CR[eg]
C[mundungus [13k?\]\: Is the komi set?
zapp [?\]\: Yes
mundungus [13k?\]\: Oh, 7.5, hold on, I need to restart.  My program must know the correct komi or it will lose by half a point.
zapp [?\]\: It is 7.5 as default
zapp [?\]\: ok
]
;B[dc]CR[dc];W[ec]CR[ec]
C[mundungus [13k?\]\: I'm going to quit and restart.
Orego [?\]\: GTP Engine for Orego (black)\: Orego (BestPlayer1/B + BestPlayer1/W) version unspecified
mundungus [13k?\]\: Did we restart the same game?
zapp [?\]\: yes
mundungus [13k?\]\: Okay, even better.  We can continue from here.
mundungus [13k?\]\: Sorry for the confusion.
zapp [?\]\: Viking is playing from its fuseki library initially
zapp [?\]\: Ok
mundungus [13k?\]\: Your opening will be much faster and better, methinks.
]
;B[cb]CR[cb]
C[zapp [?\]\: That was a unique new move!
];W[gg]CR[gg];B[dd]CR[dd]
C[zapp [?\]\: Viking is now searching 3 ply
];W[gd]CR[gd]
C[zapp [?\]\: But found G6 the best on all plies
mundungus [13k?\]\: Orego is a Monte Carlo program, similar to Crazy Stone.
zapp [?\]\: I know that much. Does it have extra go knowledge in the simulations?
]
;B[df]CR[df]
C[mundungus [13k?\]\: Not really.  Just "don't play in your own eyes".
]
;W[fb]CR[fb]
C[mundungus [13k?\]\: On a scale of 1 (black wins) to -1 (white wins), Orego estimates the value of the current game at 0.4
]
;B[fe]CR[fe];W[dh]CR[dh]
C[zapp [?\]\: 508 for viking winning -1000 to 1000
mundungus [13k?\]\: 1000 means viking wins?
zapp [?\]\: yes
mundungus [13k?\]\: Orego's down to 0.18
]
;B[ge]CR[ge]
C[zapp [?\]\: It appears that Orego favors having one solid group over two groups
zapp [?\]\: Viking played like that too
mundungus [13k?\]\: 0.26
mundungus [13k?\]\: Yes, it likes to be connected.  JacquelineGo really likes to do that, too.
]
;W[hd]CR[hd]
C[zapp [?\]\: In the end of Crazystone paper there are some good stuff to improve the simulations with
zapp [?\]\: Viking uses such go knowledge but to a larger extent
]
;B[he]CR[he]
C[mundungus [13k?\]\: What is the basic structure of Viking?
zapp [?\]\: It is alfabeta search with MC evaluation
mundungus [13k?\]\: How long has it been using MC?
zapp [?\]\: It started as Crazystone type program
]
;W[hb]CR[hb]
C[zapp [?\]\: But my search became unstable when it got more time
mundungus [13k?\]\: Are the MC games purely random, or do they include heuristics?
zapp [?\]\: They include a lot of heuristics
mundungus [13k?\]\: How many MC games do you complete per turn?
zapp [?\]\: Viking has probably the most nonrandom simulation of all MC programs
mundungus [13k?\]\: \:-)
]
;B[ie]CR[ie]
C[mundungus [13k?\]\: Ooh, blunder.
zapp [?\]\: But it is really messy code now
zapp [?\]\: And I do not know which code actually is good or bad
mundungus [13k?\]\: What language?  How many games / turn?
]
;W[cg]CR[cg]
C[zapp [?\]\: Delphi (object Pascal)
mundungus [13k?\]\: Orego is in Java and does 100,000 MC runs (from root) per turn.
zapp [?\]\: It makes 852 Simulations a second
zapp [?\]\: And evals 4.6 positions a second
]
;B[ce]CR[ce]
C[mundungus [13k?\]\: Well, that move was a bit meek.
mundungus [13k?\]\: Orego still says 0.332.
mundungus [13k?\]\: What does viking think?
zapp [?\]\: 740 which means Viking is completely sure of victory
]
;W[id]CR[id]
C[mundungus [13k?\]\: I fear Viking's opinion is closer to correct.
zapp [?\]\: Yes Orego thinks the white stones are too losely connected and can often be killed in the simulations
mundungus [13k?\]\: That may well be it.
]
;B[fd]CR[fd]
C[mundungus [13k?\]\: Orego could get away with h2, right?
];W[fc]CR[fc]
C[zapp [?\]\: Yes, this actually a tesuji Viking expects Orego to play
mundungus [13k?\]\: Are you maximizing score or probability of winning?
zapp [?\]\: Prob of winning
mundungus [13k?\]\: Same here.
zapp [?\]\: Once you try you cannot go back \:)
]
;B[bf]CR[bf]
C[mundungus [13k?\]\: 0.293
zapp [?\]\: I guess the next move by viking will lower your score because it will be ultra safe
]
;W[fh]CR[fh]
C[mundungus [13k?\]\: 0.275
zapp [?\]\: I was wrong.
];B[gc]CR[gc]
C[zapp [?\]\: But perhaps Orego still all the diagonal connections as weak
mundungus [13k?\]\: Orego often seems to make pointless forcing moves like this.
]
;W[gb]CR[gb]
C[mundungus [13k?\]\: Maybe it sees them as good because they push the evaluation farther down the tree, so it can't see that it's going to lose.
mundungus [13k?\]\: 0.426!
zapp [?\]\: It may also be the horizon effect
mundungus [13k?\]\: Effectively, yes. There is no horizon per se, but there is ... fog.
]
;B[dg]CR[dg]
C[zapp [?\]\: Viking simulations handles basic tactics well
];W[bg]CR[bg]
C[zapp [?\]\: I guess in Oregos simulations the white stones at J6 are often captured
mundungus [13k?\]\: 0.188
mundungus [13k?\]\: Time for Orego to get clever.  \:-)
zapp [?\]\: This might explain why it found G7 effective
zapp [?\]\: Had white replied at H7 the score would not have risen maybee
]
;B[bh]CR[bh]
C[mundungus [13k?\]\: Nice!
mundungus [13k?\]\: Doesn't quite work, but nice.
]
;W[ch]CR[ch]
C[zapp [?\]\: Yes, this what I like most with MC go
zapp [?\]\: Sometimes there are some really nice things happening
]
;B[eb]CR[eb]
C[zapp [?\]\: Viking now searches 4 ply
mundungus [13k?\]\: It increases as the number of moves decreases?
zapp [?\]\: Yes
zapp [?\]\: Also more moves are so bad so they need less simulations to cut off in search
]
;W[af]CR[af]
C[zapp [?\]\: It also depends on the position. Last move was 3 ply
];B[db]
CR[db]
C[zapp [?\]\: Before the move at C2 was obvious so all effort went into that variation
]
;W[fa]CR[fa]
C[mundungus [13k?\]\: Yeah, that's odd.
zapp [?\]\: Now Viking plays faster because eval is more than 900
mundungus [13k?\]\: Now 0.069 -- Orego knows it's in trouble.
]
;B[fg]CR[fg]
C[mundungus [13k?\]\: What kind of nonsense is that?
];W[hh]CR[hh]
C[zapp [?\]\: I think Orego creates cutting points, or tried to push the horizon with a small sacrifice
]
;B[cf]CR[cf]
C[zapp [?\]\: Also if it cannot find a winning move it will start to play almost random
zapp [?\]\: Thats my experience of viking
]
;W[ef]CR[ef]
C[mundungus [13k?\]\: Same here.
mundungus [13k?\]\: -0.023, Orego now knows it is behind.
zapp [?\]\: I implemented resignation to save me some embarrassment
]
;B[ig]CR[ig]
C[zapp [?\]\: Still it plays weird when it is winning as well but that is difficult to hide.
mundungus [13k?\]\: Same here.
]
;W[ff]CR[ff]
C[zapp [?\]\: I have a trick were I give more weight to points that are not clearly white or black.
zapp [?\]\: So it plays a more normal end game
mundungus [13k?\]\: Yes, I think Coulom called this the territory heuristic.
mundungus [13k?\]\: We'll probably implement that at some point.
]
;B[ci]CR[ci]
C[zapp [?\]\: Yes that is useful.
zapp [?\]\: For me it also helps move ordering
zapp [?\]\: Which is very important for alphabeta search
]
;W[di]CR[di];B[bi]CR[bi];W[ad]CR[ad]
C[mundungus [13k?\]\: We're thinking of switching to C++ for speed.
];B[hc]
CR[hc];W[ic]CR[ic]
C[zapp [?\]\: I have a new version written from scratch called Valkyria
mundungus [13k?\]\: That dropped us from -0.28 o -0.52
zapp [?\]\: But it has no heurstic yet
]
;B[bb]CR[bb]
C[zapp [?\]\: Yes. Now the white stones have more liberties.
];W[ea]CR[ea]
C[mundungus [13k?\]\: Has Viking played against Crazy Stone?
zapp [?\]\: For viking those stones were unconditionally alive in the simulations
zapp [?\]\: I am running a test with 60s thinking time right now on my laptop home
]
;B[gf]CR[gf]
C[zapp [?\]\: I think the programs are close in strength with the same harware
]
;W[ae]CR[ae]
C[zapp [?\]\: Both programs search very efftively and gain a lot from more CPU
mundungus [13k?\]\: *nod*
zapp [?\]\: Also Viking is slow for what it does.
zapp [?\]\: Thats why I rewrote Valkyria from scratch.
]
;B[ib]CR[ib];W[hg]CR[hg]
C[mundungus [13k?\]\: Orego is now playing the "Look, a monkey!" strategy, hoping Viking isn't paying attention.
zapp [?\]\: But this exactly what Viking expects...
]
;B[hc]CR[hc];W[gc]CR[gc]
C[mundungus [13k?\]\: Orego still gives -0.45. Not sure why it's so high.
]
;B[fi]CR[fi]
C[zapp [?\]\: E3 has only two liberties
];W[gi]CR[gi]
C[zapp [?\]\: So by chance where is a 25% chance that black plays both
zapp [?\]\: And then black wins
mundungus [13k?\]\: Yes, or slightly less if white get e1 first.
mundungus [13k?\]\: Noe -0.71
mundungus [13k?\]\: /Now
]
;B[ha]CR[ha];W[ih]CR[ih];B[ed]CR[ed];W[bc]CR[bc]
C[mundungus [13k?\]\: Orego thought b7 was a really bad move\:  -0.31
zapp [?\]\: Note that the upper group can also be killed in completely random simulation
]
;B[eh]CR[eh]
C[mundungus [13k?\]\: Yes.
];W[fg]CR[fg]
C[mundungus [13k?\]\: Orego was expecting e1, and may try a3 next.
];B[da]
CR[da];W[ab]CR[ab]
C[mundungus [13k?\]\: Some sort of tactical decomposition is in order.
]
;B[ga]CR[ga];W[ia]CR[ia]
C[mundungus [13k?\]\: That moved us from -0.35 to -0.72
zapp [?\]\: I fear we might get a superko problem in the upper right
mundungus [13k?\]\: Does Viking handle superko?
zapp [?\]\: Yes, now it finally does it correctly.
mundungus [13k?\]\: Hmm, something very strange is happening.
zapp [?\]\: It just situational in stead of positional super ko for some reason I do not remember.
mundungus [13k?\]\: I think Orego may have crashed.
mundungus [13k?\]\: The processor load just dropped precipitously.
mundungus [13k?\]\: Let me try restarting...
zapp [?\]\: Ok
Orego [?\]\: GTP Engine for Orego (black)\: Orego (BestPlayer1/B + BestPlayer1/W) version unspecified
mundungus [13k?\]\: Okay, now we're crunching numbers again.
mundungus [13k?\]\: Hmm, dropped off again.
mundungus [13k?\]\: Well, in any case, it's clear that Viking has won this game.  I'
mundungus [13k?\]\: ll resign on Orego's behalf.
mundungus [13k?\]\: Good game, and congratulations.
zapp [?\]\: Ok, thanks for the game!
mundungus [13k?\]\: As the winner, would you like to report the result to the ladder?
zapp [?\]\: I could do so so.
mundungus [13k?\]\: Thanks.  I find it somewhat satisfying that the first program able to stop Orego was another MC program. It seems a very promising line of research.
zapp [?\]\: Yes, I was amazed by the progress I did in a few months
mundungus [13k?\]\: Are you participating in the KGS tournament this weekend?
zapp [?\]\: Still making it work well on 19x19 is hard work
mundungus [13k?\]\: (yes)
zapp [?\]\: Probably
mundungus [13k?\]\: Okay, I'll see you there.
zapp [?\]\: See you
Xrillo [-\]\: börjat jobba med viking igen?
]
)