Closed danzel closed 6 years ago
Should also do an alternative one where we late evaluate where pieces are placed. Don't place them as we go, but instead when enumerating possible future moves (Expand) ensure that we could place the next piece if we had placed all the previous ones.
Added one, current results (It is B/1):
Run 0
Starter 0, winner 0
Starter 1, winner 1
Draw
Run 1
Starter 0, winner 0
Starter 1, winner 1
Draw
Run 2
Starter 0, winner 0
Starter 1, winner 0
^^^ Not Equal ^^^ A Winner: True
Run 3
Starter 0, winner 0
Starter 1, winner 0
^^^ Not Equal ^^^ A Winner: True
Run 4
Starter 0, winner 0
Starter 1, winner 0
^^^ Not Equal ^^^ A Winner: True
Total Time 11520 / 73446
Probably need to give it (at least) twice as many iterations as it takes up an iteration to place pieces.
Ran MOMCTS+1_6 (10000) vs MCTS(20000) https://gist.github.com/danzel/701c4189575fa136e58711e690ba3385 Wins go: 14 v 8
When both are given 100k iterations, MCTS wins (7 wins vs 19), but it takes 7x longer. When running for equivalent time MO is much stronger (within any time range I've tested).
Maybe at really high iterations MCTS will be stronger?
Results running 700k vs 100k:
0 Z ^^^ Not Equal ^^^ A Winner: False
1 Z Draw
10 Z ^^^ Not Equal ^^^ A Winner: False
11 Z ^^^ Not Equal ^^^ A Winner: True
2 Z ^^^ Not Equal ^^^ A Winner: False
3 Z ^^^ Not Equal ^^^ A Winner: True
4 Z ^^^ Not Equal ^^^ A Winner: False
5 Z ^^^ Not Equal ^^^ A Winner: True
6 Z ^^^ Not Equal ^^^ A Winner: True
7 Z ^^^ Not Equal ^^^ A Winner: True
8 Z Draw
9 Z ^^^ Not Equal ^^^ A Winner: True
Total Time 2802051 / 2598379
MOMCTS is doing slightly better, (1 more win), but it also had more time.
Current one thinks it can buy pieces that it cannot place (especially later on). If MCTS cared about placing pieces then it would be stronger.