- Created Wednesday, March 23rd 2016 @ 21:31:48
I think the discussion is worth an topic on its own. Therefore this topic.
We had so far a lot of 'brute force' challenges (as pointed out by others). Those are fun on there own, but it writing a good alpha-beta search tree, and a smart bit representation of the state is key.
The nice thing about Go is (and that is the reason that I proposed it), is that such mechanisms don't work. You have to come up with something totally different. That this will lead to bots that are weaker than human players, and therefore play a bit dumb should not matter. Every participant has the same handicap, so an equal playing field. As the game is 19x19, I guess that is the preferable board size.
The only reason I can think of, to scale down, is to make games take less long, but to overcome that I would rather propose less thinking time per move. 100 ms for instance. Yes, that will make bots even weaker, but again, its all about an equal playing field.
- Created Thursday, March 24th 2016 @ 01:44:53
I agree completely. A full size board with no turn limit makes for an interesting challenge, and this will remain true with shorter time limits.
I think 200ms would be more interesting than 100ms, but I would still prefer a very short time limit instead of a smaller board or a turn limit.
- Created Thursday, March 24th 2016 @ 02:42:42
200ms and no move limit will be great (even if 100ms)
- Created Thursday, March 24th 2016 @ 07:25:44
I also like 19x19 with 100 ms (or200) and a more generous limit. But there must be some limit, because most bots wouldnt know how to pass, games can become extremely long. 200 moves should be enough?
- Created Thursday, March 24th 2016 @ 10:30:51
The problem with imposing any limit on the number of turns is that it fundamentally changes the nature of the game: guarding against the possibility that your opponent will "bomb" a secure area at the end of the turn limit becomes part of the strategy. Searching for good areas to bomb also becomes part of the strategy.
The resulting game is distinctly different from Go. Even if the turn limit is high enough to allow decent bots to play Computer Go without being cut off, that is no longer the game they're playing - the rules have changed in a manner which allows you to create living groups in areas which cannot support them.
I'm strongly in favour of the full size board - this challenge will benefit from domain specific knowledge, and there's a lot of knowledge out there which has grown out of the 19x19 board. However, I think removing the turn limit is even more important. In order to have both, I'd be happy to see the time per move made however short it needs to be. If we can't have both, even with a very short time limit, I'd compromise on the board size.
- Created Thursday, March 24th 2016 @ 15:24:29
I like Jorose's proposal: "if you capture a certain number of stones you win!". 50-100 stones should be plenty. Really, without a limit, games between bad bots could go on for thousands of moves. I did a test and let my program play against itself with a limit of 10 000 moves and this limit was easily reached.
- Created Thursday, March 24th 2016 @ 16:56:46
Okay, maybe I was wrong and a turn limit is needed.
If that's the case, I hope it can be set high enough that bots will always have enough moves to fill their territories with a checkerboard pattern so that the nature of the game is not changed.
- Updated Thursday, March 24th 2016 @ 17:40:32
Make limit of 400 moves for example. There can be bad bots, but why good ones must suffer? Limitation kills strategic component of playing. Remains only tactics. Better board representation, good optimisations.. and so on
- Created Thursday, March 24th 2016 @ 20:48:06
I agree with higher round limit (say 300) and if necessary lower time limit. Possibly in combination with a score limit: if you have N points, then you've won, with N = say 200.
But let's start with making some good bots.
We have to hack!
- Updated Thursday, March 24th 2016 @ 21:18:05
There will be some bad bots. That's why i see this solution best: calc score by GnuGo or by another program at 300th move
- Updated Saturday, March 26th 2016 @ 13:15:17
Based on further local testing, I think 19x19 should be possible if we can have the round limit set to 250, or maybe 300 at the most (meaning 500 or 600 turns).
If bots don't fill their own eyes, they should be able to have the board in a state where it can be scored correctly and unambiguously at that point. If some points are still in contention after 600 moves, one or both bots is (probably) being silly and it's okay to cut off the game.
Extrapolating from these tests, here are some suggestions for different board sizes and their associated turn limits and maximum time per game if the engine is instaneous, first at 500ms/move and then at 200ms/move:
19x19: 300 rounds (600 moves), 5 minutes per game at 500ms/move, 2 minutes per game at 200ms/move 17x17: 240 rounds (480 moves), 4 minutes per game at 500ms/move, 1.6 minutes per game at 200ms/move 15x15: 187 rounds (374 moves), 3.2 minutes per game at 500ms/move, 1.25 minutes per game at 200ms/move
In reality the engine may take a nontrivial amount of time (detecting long ko cycles in long games), so actual game processing times would be longer, maybe even double or more.
- Created Monday, March 28th 2016 @ 04:15:06
How about giving each bot 2 minutes time bank at the start of the game, but don't add additional time for every move? This would limit the length of each game at 4 minutes, regardless of the number of moves in the game. But bots would need to keep track of their time carefully in this case...
- Created Monday, March 28th 2016 @ 16:04:57
Great idea, btzy! I agree