Playing against a board game L'Oaf AI

After publishing the original post, I noticed an edge case where my implementation drifted from the rules. If a round’s effect targeted the highest or lowest reputation player and both players were tied, my code applied the effect to only one player. The rules (and the intent of the condition) are that tied players should both be targeted.

This does not flip most games on its own, but it can nudge optimal play and it definitely matters when you are training and comparing strategies. I fixed the targeting rule, retrained the models, and reran the same analysis scripts. By the way: 8 is now the least favorite card, instead of 9.

A risky bot mode

While building the playable version, I ran into a practical issue: the strongest bot is good in a way that feels a bit boring. When the position gets close, it often goes into mutual failure where both players end up fired, which produces a draw. This is of course a good strategy if your objective is to avoid losing, but it is not the most fun opponent.

So I trained a second version of the 128×128 model with a deliberately skewed reward: win +2, draw −1, loss −1. This makes draws feel as bad as losses, which pushes the policy toward trying to win instead of forcing a draw. You can switch between the normal and risky model in the game UI. Below are some plots of this new model, and it performs as expected: it is a bit worse than the all-rounder models, but it's still good. It learns at approximately the same pace and generally will still beat humans, I think. What's interesting is that it uses different moves than the other models, using high cards more often.