Ancient Roman Board Game
ludus-coriovalli.web.app
ludus-coriovalli.web.app
> Using Alpha-Beta search agents — the same class of algorithm that powered early chess computers — the team ran 1,000 simulated rounds for each candidate ruleset, allowing one second of processing time per move. The AI tracked which lines on the board were used most frequently during play, generating detailed edge-usage statistics....
> Nine game configurations matched the wear criteria. All of them were blocking games, and the most frequently matching format was a four-versus-two game in which pieces start on the board. This site faithfully reproduces one of these AI-validated configurations.
I would say this is more "inspired by" Ancient Rome.
I imagine the incentives of having a crisp story for media consumption don’t help. I’d hope to read a lot more: “we’re missing the majority of the pieces to this puzzle. This represents our best guess given current evidence and methods.”
Eg much is not known about dinosaurs. Many things cannot be found in fossils (obligatory xkcd: https://xkcd.com/1747/)
How do you communicate what is unknown or what can't be known?
We tend to just drape a bounding skin volume over the bones and give it a color, MAYBE some feathers.
But a whale skeleton looks nothing like the mass of flesh of a whale. Skeletons give no clue about an Elephant’s head shape.
So imagine a huge fluffy owl or yellow chick with a trex skeleton deep inside.
That being said, "We asked an AI..." is a special kind of uncertainty that goes above and beyond anything else Archaeologists do.
--------
"No written rules for this game survived antiquity. To reconstruct how the game may have been played, researchers turned to the Ludii General Game System — a comprehensive digital platform developed at Maastricht University that can model and simulate thousands of historic board games. The results were published in the journal Antiquity (Volume 100, Issue 409, 2025).
Using Alpha-Beta search agents — the same class of algorithm that powered early chess computers — the team ran 1,000 simulated rounds for each candidate ruleset, allowing one second of processing time per move. The AI tracked which lines on the board were used most frequently during play, generating detailed edge-usage statistics.
These statistics were then compared to the physical wear patterns on Object 04433. To account for human cognitive biases — such as right-handed players preferring to play on the right side of the board — the researchers applied symmetry transformations to the simulation results, maximising consistency between AI-generated play and the actual marks left by ancient players.
Nine game configurations matched the wear criteria. All of them were blocking games, and the most frequently matching format was a four-versus-two game in which pieces start on the board. This site faithfully reproduces one of these AI-validated configurations."
--------
It's interesting that they considered use-wear on found pieces as input for their AI. Still, this study made a lot of assumptions. I wouldn't be surprised if a different team could use the same methods and come up with a completely different result.
Also, they didn’t “ask” an AI anything, as it was not a natural language AI tool.
So they could have used wear patterns as a component in an evaluation function used in minimax but I get the feeling they most likely did a wear analysis afterwards to weed out some of the possible rulesets indicated by the minimax play patterns.
______________
[1] https://en.wikipedia.org/wiki/Alpha%E2%80%93beta_pruning
Your complaint is like complaining that a CS paper doesn't even mention that P = NP is still unknown, and that it just assumes that the best sorting methods are O(n log n). Some things are considered general knowledge in a field, and in archaeology one of those things is: all of this is a best guess given current evidence and methods.
The whole paper is full of citations to other work, which is full of citations to other work - all this work linked includes detailed reports of what was found where and what-else was there, people who do statistical analyisis of similar findings, people who ask questions like "what explanations can there be for it?", more importantly "what else would we find if X was true, if Y was true?". When new evidence arrives, people can and do go back and re-examine these things. Your random skepticism and all the questions you may ask have already been asked and addressed, and frankly: these archaeologist's conclusions carry far far far more believable weight than your half-assed skepticism.
Model and simulate based on what?
> Nine game configurations matched the wear criteria.
So their idea was to generate candidate rulesets, have AI try to figure out rational play, then see which pieces would be moved most often and match that to the forensic evidence?
However it would still be useful if archeologists used the board to figure out some games similar to checkers, or go; or if they also have the pieces they could guess it was a combat game like Shogi. Any of those would give you insight about the kinds of leisure that people may get from that board.
Even if the researchers did not uncover the exact right set of rules, it would probably not be dissimilar from an actual variant that could have been played somewhere in Rome.
Here's a tablebase analysis for a simple bear game I constructed a while back: https://emarzion.github.io/coqtbgen/
I was thinking of them yesterday because I noticed one for sale in the antiquities shop in Andor, which brought up all sorts of Earth/Rome/Star Wars cannon questions.
The segment where she uses a replica dodecahedron to make a chain from aluminum wire (stiffer than gold, meaning this would have been easier with actual gold) is entirely convincing. It explains everything about the shape of the objects, both the protruding knobs on the end and also the fact that the holes are of different sizes.
https://www.academia.edu/26809664/The_enigma_of_the_dodecahe...
A) make sure you get at least one hare to one of the central two spots (will always be possible) and camp it there. Keep your other hare moving around on your side.
B) the opponent will have to bring two hounds over to catch your other hare. To do this they will have to create a gap between the first and second hound in the middle spot above/below your camped hare. As soon as this happens, move your camped hare upwards.
It's pretty easy after that TBH. The AI has usual burnt a load of moves by that point and you have so many options from that position
As hares, if you rush one of the middle nodes, the hound machine could capture, but it never figured out moving two hounds all the way over to do so in my play. As hounds, it's basically the same, you just make sure none of the hares ever get "behind" you because then it's a lot harder to contain them.
Edit: Or, put a different way: Sometimes the rules to games are lost for a reason; they're not very good.
Are there other known ancient games that work like this?
Similar in that it is asymmetrical, very old, and we don't quite know what the rules were (though with tafl we have slightly better historical evidence, including an account from Carl Linnaeus in 1732, which has allowed us to produce a few educated guesses). In fact tafl is sometimes speculated to derive from the Roman game of ludus latrunculorum - I'm not sure if that is the same game described here.
Is it like chess only way more obvious, that the first mover wins, and if so, wouldn't the Romans have figured it out?
Or am I just exploiting a weakness in the AI and a human would make better choices?
There needs to be something to stop a deadlock. Like an element of luck for the hares?
The videos on the game (and all his other videos) with Irving Finkel, a curator at the British museum, are spellbinding. He has the looks, manners and enthusiasm of an eccentric museum curator from central casting!
This is much more palatable and cool since they’re not just randomly guessing what the game can be like this article is
They are creating a probability distribution function where each point is a different gameplay ruleset and then we are free to pick the gameplay ruleset with the highest probability, with the caveat that there are other similar gameplays, with slightly different rules, and almost the same probability. And there are other gameplay rules with much lower probability because they don't match the wear of the board.
A random guess would have a flat distribution where any gameplay has the same probability.
figured out that my hares needed to stay together in either the top or bottom center and left corners of the big square.
at one point the bot started wasting moves so i repeated mine until it ran out! not sure if I found the deterministic win of the game or it was a bug in the AI.