ZeusAI paper: Learning to Play 7 Wonders Duel Without Human Supervision

[giovep]

We wrote a paper on ZeusAI and posted it on arXiv today: https://arxiv.org/pdf/2406.00741. It will be published as a short paper at the Conference on Games 2024!

We outlined how ZeusAI works, as well as some of our findings and proposals to reduce the 1st player advantage (which we estimate to be around 66.8%). Here is how victory types and wonder preference evolved as ZeusAI became more and more proficient at 7WD:




Check out our paper for more details! Also, watch our past ZeusAI live streams and follow the YouTube channel to know about future 7WD live streams and videos!

The ZeusAI team (giovep, sirion, Ossido, Sashair)

[AutumnsTale]

Thank you for the paper! I have some thoughts below.

I'm an amateur Go player and The Castles of Burgundy (CoB) player. I spent about 3 months spare time on CoB AI and finished an endgame (Phase E4, E5) solver. After that I tried to make a handcrafted evaluation function, but it performs really bad as expected. I don't have enough engineering time and GPUs, so the project was stopped. AlphaGo showed lots of amazing strategies. I believe this may happen on board games. Thanks to ZeusAI for making this possible.

1. CoB as a popular game is similar to 7WD, no hidden info, with randomness. CoB has a player hex duchy which is a graph like structure. I believe this mechanism makes NN much larger. Are you and your team interested in making CoB AI?

2. game balance:

2.1 Balance in board game is a big problem. CoB top players sometimes have 90% confidence to win the game right after the first draw in phase A (no action at all) if the draw and dice numbers favors to him/her a lot.

2.2 Forget about the costs, AI may help game designer and tester to make more balanced games. More metrics may be developed other than win rate.

Based on my experience in Go, 50-50 win rate doesn't neccessarily mean balance. In complex situations, although Go Engine indicates an even game, the error tolerance may differ to 2 players. For example, Aristotle has 10 possible choices to keep the value, but Cleopatra has only 1 counterintuitive choice. The balance for human players is not only the win rate, but also other metrics that may be distilled from the MCTS. I heard that some Go Engines have this "complexity" metric and use this in handicap games.

3. What do you think of board game AI of game with hidden info? I suspect it would be totally another story. Suphx (Mahjong AI) looks very different with AI of perfect info games.

[giovep]

1. CoB as a popular game is similar to 7WD, no hidden info, with randomness. CoB has a player hex duchy which is a graph like structure. I believe this mechanism makes NN much larger. Are you and your team interested in making CoB AI?

In an ideal world, yes, but I don't see myself having enough time for this...

More metrics may be developed other than win rate.

I agree, but this becomes much more complex than estimating the win rate.

If we focus on the 1st vs. 2nd player balance, I would say that we mostly care about: (1) the win rate (equivalently, the true value of the initial position before any setup); (2) the variance of the true value at some point early in the game (e.g., after the setup and before the first move, or after the wonder selection phase). Specifically, we care about the win rate being close to 50% and about the variance being close to 0 (the reason being, as you pointed out, that it is unpleasant to have a 90% value in an early position).

We did not mention the variance in our paper (we had a 4-page limit), but it looks like our wonder selection variants produce a lower variance, compared to the standard rules. So that's also good.

3. What do you think of board game AI of game with hidden info? I suspect it would be totally another story. Suphx (Mahjong AI) looks very different with AI of perfect info games.

Yes, it is a completely different story! It opens up the whole realm of inferring the unknown information from the known information, and also bluffing (if the game is not cooperative). Hanabi and Bridge are two games that I would like to think about, if I had the time.

[AutumnsTale]

Hanabi and Bridge are two games that I would like to think about, if I had the time.

I think bridge is difficult since it doesn't have a standard process. The info in bidding is vague. In a bridge match, players can ask opponents to explain the info behind the bidding. I don't know how programmers and bridge players can do the test

[Bedr2341]

Very interesting paper, thank you for sharing . Interesting to see how the Great Lighthouse is less preferred by the AI

[enjoyment_enthusiast]

Great work

It really affected my valuation of Tavern. Now I pick it whenever possible with few exceptions like resource monopoly or science supremacy
Previously I often picked resources over Tavern

It would be nice to have it built in into BGA one day

[SYushen]

Can we have a longer list of most preferred cards?

[giovep]

Can we have a longer list of most preferred cards?

Sooner or later, yes! After the Summer vacation, the plan is to work on more in-depth statistics on ZeusAI's games.

[ifasoldt]

Very interesting stuff! I'm a pretty strong player (generally in the top 150 or so) and these findings correspond very strongly with my own understanding if the game, with perhaps a small exception being that as I've played more, I've tended to start preferring Piraeus where before I preferred the Temple of Artemis. I wonder if this is because I play a wide variety of player level-- Piraeus is a more conservative wonder-- perhaps I find that as long as I can guarantee grey resources using it, I can rely on outplaying the other players over the course of the game. That is, as the generally high rated player, I try to minimize variance.

The first player advantage also lines up with my sense-- I've said before that I think the advantage is worth at least as much as having one more extra turn wonder (people strongly disagreed when I posted this), but it seems the AI thinks it's even stronger than this. Going first is essentially as strong as getting to pick the 4 wonders you want, which is amazing.