So I did something this weekend and thought I would share. I made an engine which can fully simulate Wingspan games which allowed me to train my own Wingspan AI. Currently, playing itself in 3 player games with all expansions it averages 100 points, which I think makes it comparable to a fairly strong human, though I had trouble finding any statistics about connecting average final scores with skill in this (or any other) configuration.
After simulating a lot of games with this bot playing itself, I generated a tier list. The basic idea is that across hundreds of thousands of simulated games we can look at every time a card was drawn blind from the deck (not the tray), and at the end of the game record how far that player finished ahead of or behind their best opponent. For example, take every time Chipping Sparrow was drawn off the deck in round 1, average those final score margins, and compare that with the same average for every other bird drawn in round 1. This final quantity is the Chipping Sparrow's round 1 value.
A problem that remains with this is that the bot refuses to basically ever play ~18% of birds. This in itself is some strong signal, but bots might not be able to properly see the value of certain birds, so in order to get a little more data, when a bot chooses to play a bird, we force it to choose a random bird instead of its intended one 2% of the time. This means the bot will sometimes play a bird in its hand it doesn't want to play, though after this, it'll still do its best. This lets us get some more information on those "bad" birds.
Huge caveat that this metric is derived from pure bot play. While I think my bot is pretty good, it's probably not stronger than the strongest human, and also, it's somewhat unrealistic to be in a game where all players are matched in skill, which could distort the perceived value of, for example, pink powers. Moreover, there might be bugs in my simulation engine. That said, I think these results look reasonable enough to me at an intuitive level that I think these findings ought to transfer pretty well to real human play. (When I play this bot personally in a 3 player game with 2 bots, I win only 50% of the time or so, and the bot subjectively feels like a strong human to me.)
Some highlight findings:
* Wood Duck is the #1 card to have in your starting hand
* White Stork is overall the most powerful bird
* In Oceania, the power four are far away from S tier: ravens are both in D tier, Franklin’s Gull in C and Killdeer in F.
* Endangered Species Protector is 100% busted (okay, I'm editorializing a bit, but come on...)
Anyway, here's the list: https://horse-feed.github.io/wingspan_tier_list.html
After simulating a lot of games with this bot playing itself, I generated a tier list. The basic idea is that across hundreds of thousands of simulated games we can look at every time a card was drawn blind from the deck (not the tray), and at the end of the game record how far that player finished ahead of or behind their best opponent. For example, take every time Chipping Sparrow was drawn off the deck in round 1, average those final score margins, and compare that with the same average for every other bird drawn in round 1. This final quantity is the Chipping Sparrow's round 1 value.
A problem that remains with this is that the bot refuses to basically ever play ~18% of birds. This in itself is some strong signal, but bots might not be able to properly see the value of certain birds, so in order to get a little more data, when a bot chooses to play a bird, we force it to choose a random bird instead of its intended one 2% of the time. This means the bot will sometimes play a bird in its hand it doesn't want to play, though after this, it'll still do its best. This lets us get some more information on those "bad" birds.
Huge caveat that this metric is derived from pure bot play. While I think my bot is pretty good, it's probably not stronger than the strongest human, and also, it's somewhat unrealistic to be in a game where all players are matched in skill, which could distort the perceived value of, for example, pink powers. Moreover, there might be bugs in my simulation engine. That said, I think these results look reasonable enough to me at an intuitive level that I think these findings ought to transfer pretty well to real human play. (When I play this bot personally in a 3 player game with 2 bots, I win only 50% of the time or so, and the bot subjectively feels like a strong human to me.)
Some highlight findings:
* Wood Duck is the #1 card to have in your starting hand
* White Stork is overall the most powerful bird
* In Oceania, the power four are far away from S tier: ravens are both in D tier, Franklin’s Gull in C and Killdeer in F.
* Endangered Species Protector is 100% busted (okay, I'm editorializing a bit, but come on...)
Anyway, here's the list: https://horse-feed.github.io/wingspan_tier_list.html