Are you only looking for other algos to improve your bot's performance? Curious if there's other routes to take to get better gains b/c it sounds like you're using pretty good algo's already
#AI in board game Pylos
wet venture
wet venture
That's super fair!
Have you thought about experimenting to figure out your own heuristics? If they're simple and effective, might be able to guide rollout
Data structures that store evals of seen states to avoid re-eval, iterative deepening, and parallelization are more general options if you haven't already tried them.
Also, qualifier: I'm throwing darts and this isn't something I'm that experienced with lol
wet venture
If the effort/challenge needed to overcome making your own heuristic makes it a less effective route, yeah, I agree. You can be systematic in your approach to help, but it's a challenge, imo. Gotta balance the effort with payoff
Not too familiar nor am I too familiar with Pylos, so I doubt I can pass on any specific thoughts here. I'm sure you already know the essentials, which is all I got (e.g. number of simulations vs. cost if that's part of your performance metrics somehow, how much it values exploration, and like you mentioned rollout is a big consideration).
You probably can w/ TD or Q-learning, tables/replay, some more manual/meta analysis around successful opponents (e.g. pattern recog, clustering recurrent tactics), etc.
Right, sorry, I keep losing track of that here
Gains, at least in concept, can help guide me to possible improvements, but given my lack of experience, you'd probably see me throwing the kitchen sink while tracking impact on how fast it moves per turn
IDK of a better way, so glhf! If you got the time, let me know what you ended up going with and why
That's good context on why you chose this strategy, thank you.
How are you tuning the explorationWeight? Curious to learn more about MCTS and how you're using it here
wet venture
Oh okay, this tracks! I was wondering if there was something neat to use like GridSearch to save my LOC, but it sounds like this can probably just be done more directly/simply