03-01-2021, 01:13 PM
Hmmm no I can't think of any heuristics we currently have that would effectively check for such patterns.
(centreProximity) could help if you say that the center is important, but that applies throughout the whole game, not only in the beginning. So if it's really only important in the beginning of the game, maybe it doesn't help.
I see that this game incrementally updates the players' scores during the game, it doesn't only compute them at the end. This means that the (score) heuristic might be useful, especially if the scores tend to gradually increase with strong play. If there are huge swings in scores throughout the game, and the score values in the middle of the game are not reliably indicative of which player is in the strongest position, it might not help...
The "features" for our Biased MCTS AIs may indeed be better for providing those kinds of geometric-pattern-based hints, but it can be even more difficult to correctly assign their weights than it is for the heuristics.
If you don't mind waiting for several weeks, I am busy automatically training and evaluating heuristics and features and different AIs for all the games we currently have, and expect results... probably somewhere between now and 1 month. I'm pretty sure Throngs is also included in there (only games that were added to Ludii very recently may be excluded), so maybe the automated training will find something useful there.
(centreProximity) could help if you say that the center is important, but that applies throughout the whole game, not only in the beginning. So if it's really only important in the beginning of the game, maybe it doesn't help.
I see that this game incrementally updates the players' scores during the game, it doesn't only compute them at the end. This means that the (score) heuristic might be useful, especially if the scores tend to gradually increase with strong play. If there are huge swings in scores throughout the game, and the score values in the middle of the game are not reliably indicative of which player is in the strongest position, it might not help...
The "features" for our Biased MCTS AIs may indeed be better for providing those kinds of geometric-pattern-based hints, but it can be even more difficult to correctly assign their weights than it is for the heuristics.
If you don't mind waiting for several weeks, I am busy automatically training and evaluating heuristics and features and different AIs for all the games we currently have, and expect results... probably somewhere between now and 1 month. I'm pretty sure Throngs is also included in there (only games that were added to Ludii very recently may be excluded), so maybe the automated training will find something useful there.