Initial big work from Gunnar Farneback: MC simulator with alpha/beta on top of GNU Go 3.7.9.
This version is simplified, and does 3500 nearly random games for each of the top moves, whose value is more than 1/3 of best value. Average on a game seems to be near 3 * 3500 = 10 500 simul.
Modif made by Alain B.
- use probability of victory instead of mean score.
- remove alpha beta
- do much less simulation for good speed on CGOS, easier understanding, and reference for future re-evolutions
- Dont call MC if the best move is SEMEAI. This maybe weakens the engine (at least on regression tests it does).
- small modif of the random game generator (knows tobi)
- trivial timing optimisations (don't call MC when only one move is considered)
- Set up another bot with no semeai restriction and compare...
- Put back alpha/beta, maybe incrementaly
- Try to make it work on larger boards.