Board Game Arena

Pickomino is now officially part of the Gymnasium external environments for Reinforcement Learning.

You can read more and train your AI model or agent here:

https://github.com/smallgig/Pickomino

It does feel like the kind of game that a NN would do well at. Perhaps NN rollouts combined with MCTS. Does this use the 40th anniversary rules?

Tempted as I am to play around with such code, I'm having too much fun just playing the game myself. I did make a MCTS bot to train against, and it easily solves the short-term problems, but it doesn't see the big picture. It just optimizes for the maximum worm reward of the turn. There's more to this game than that.

A NN could learn bigger-picture rules like when to play risky, and when to play for the steal, and evaluate just how good having that 4-worm tile really is when there's a 20% chance you lose it next turn. I wonder what sorts of environment variables it would find most useful.

Feel free to try out the environment. We would love to get some feedback on it.

Does the observation space need the stack height of all players?

Does the observation space need to a flag for when a tile was stolen? To make reward shaping easier?

But basically, does the environment work? Can you train a model and what are the results?

Thanks!

What are the differences in the anniversary rules?