I built a Repository to Test Hanabi Strategies

[Jennifish]

Hi All,

I wound up doing some research on Hanabi, and built an engine in python to test Hanabi strategies. If you have any interest in it, the github repo is available here github.com/Jericgordon/Hanabi-Math. The conclusions to that research mostly involved potential maximum scores. My research is also in the repo.

Happy to answer any questions about it.
Jen

[Jennifish]

For anyone interested in results, There are 2 large ones.

First, Hanabi is almost always winnable with 2 players. Below is a graph of Hanabi scores if you cheat. The average works out to a score of around 24.



Second, The greatest factor we could find to score was clue efficiency. To try to test strategies in general, we implemented "cheating" but with a cost; that every played card would cost some fraction of a clue, between 0-2 clues.

Basically, if you wind up with a clue efficiency of 1 per card, you'll have a mode score of 21 in a 5 color game, but the full graph is below

[_mg6maciej_]

Is the first graph showing 2 player 5 color hanabi game cheating stats? Is that hand cheating or deck cheating?
It seems to suggest over 10% of hanabi games cannot be won, which is way too much if that's the variant you were analysing.
According to the research of a group of people who are active on hanab.live, the cheating winrate is >98.5% in 2p 5 color game. Btw. for 5p 5 color cheating winrate is way higher at 99.98%.
If you want, contact me via private message, and I'll give you contact to that group, so you may exchange your observations.

[Jennifish]

That is what my findings say at the moment. I will qualify that this is without flamboyants, and without knowing the deck in advance. You could get a higher score if you knew the deck in advance, but as I'm interested in the best score a strategy could get, that number doesn't interest me at all

https://github.com/Jericgordon/Hanabi-M ... heating.py

I set up this to run the game, and print everything, and stop when the strategy I currently think is optimal doesn't get a perfect score on a game. If folks would like to look at that, I'm happy to be wrong, and work with people on developing what is optimal.

[_mg6maciej_]

Ahh, okay. So that's hand cheating. The numbers can be this low then I guess, because deck cheating allows you to optimize a lot more, knowing when you draw the other copy of a card.
But it's also a lot harder to figure out exact hand cheating winrate, because of all the different strategies you might take like trying or not trying to keep useful 1-aways, stalling in situations when a card in your hand will play in a few rounds, etc.
I'm personally not very interested in either hand or deck cheating winrate, and I'm more interested in stats for 4-5p, maybe 3p too, but not 2p.