Open arr28 opened 9 years ago
Looking through the logs...
This game must be a really good exercise of the hyper-edge code. I wonder if that's where the problems are here.
The whole game only consisted of 46 "pick" moves and from the point where I began the commentary, it was only 26 moves long. Maximum branching factor was 6 and average branching factor was probably more like 4. That's still 4x1015 moves though.
Thinking about this game, it's probably a real shocker for MCTS. Random moves are a really bad approximation to reality. Until there is significant branch completion happening, the values coming back from MCTS must be extremely noisy.
I suspect there's no "bug" here and that it's just a game that sucks with MCTS. Re-tagging accordingly.
This game displays a similar issue: http://www.ggp.org/view/tiltyard/matches/31da6caf28a7b3ed393e864747c553b6758d1320/
Around turn 14 or so, in a forced move sequence, the estimated score drops rapidly even though the tree is utterly unbranching in this time.
Interestingly this seems to happen about when the node table fills up
Consider whether this is also related to #229 where there was suddenly a huge drop (to 0 in this case) in D&B. These are both games with a considerable "1 player retains control for a while" element so consider issues with hyper-edges.
We lost this match of Kalaha 6x2x4. Local logs (not uploaded because they're too large) showed an expected score of ~80 up to about half way through the game. Then came a "surprising move" (virtually all of the node pool freed, score immediately drops to 50). A little while later, our expected score dropped to 0 and remained there.
I note that GreenShell's logs are the inverse of ours. Its score decreased down to <20 then suddenly jumped up 50. So the surprising move surprised both of us.