Fall 2026brianmars.comBuilt at EGBT Technologies

How long does it take to learn rules nobody tells you?

I simulated players learning a card game that never explains its scoring. The best of them needed 58 rounds. I'm 58 and still learning the other game.

I'm building a small card game as part of a side project. It keeps my skills sharp, and it's making me learn some areas of design I haven't worked in before. I'm not ready to say what it is yet, so I've blurred the details here on purpose.

I wanted to build the logic for the computer player and after a little research I found the expectimax algorithm that was fast enough to run on every hand. I ran a few thousand hands and it was good. But it wasn’t good enough to overcome playing first. Playing first, over the course of a round, almost always led to winning. The game I am modeling is from an older game and in it the player (you) goes first. But you rarely win… at first.

The piece I want to talk about today is the scoring, because the game never explains it. You are not told which cards go together, how many points a pairing scores, or which pairing beats which. You play a round, and a screen at the end shows both hands, names what each one made, and gives the winner's points. Everything you know about the rules comes from those screens.

I wanted to know how long it takes to learn the rules. So I built an expert player that knows the real scoring, and then some learners that don't. A learner starts with almost nothing, and after each round it reads the screen and updates its beliefs. Then I counted how often a learner's move was as good as the expert's in the same spot.1

It's also a fair description of how I, an autistic human, get through a conversation.

Two ways to learn

There are two kinds of learner.

The reasoner chains its sightings together. If it has watched card A score with card B, and B score with C, it concludes that A and C go together too. The recorder only knows the exact combinations it has seen with its own eyes, and it never guesses past them.

Give both of them perfect memory and the reasoner matches the expert on 90% of its moves from round 58 on. That's about three full games. The recorder is at 84% after 100 rounds, and it still hadn't reached 90% when I stopped it at 150. A player who never learns anything matches the expert about 20% of the time.

Fig. Perfect memory, 1,000 simulated learners of each kind. The reasoner stays above 90% from round 58.

Scores and rankings come quickly, because the screen states them outright. Which cards go together is the slow part, since a round shows two pairings at most and plenty of rounds show none.

Then I let them forget

People do not have perfect memory, so I added forgetting. Each fact fades with the number of rounds since the learner last saw it. Seeing it again resets the clock and makes it last longer.2

The forgetful learners stop improving partway, because they lose facts about as fast as new rounds remind them. Where they level off depends on how fast they forget, and playing longer doesn't move it. With fast forgetting, the reasoner settles at 79% of expert moves and the recorder at 63%.

Fig. Average of the last 20 of 150 rounds, 500 simulated learners of each kind. A player who never learns matches the expert 21% of the time.

The reasoner holds up better because it can rebuild a forgotten pairing from others it still remembers, while the recorder has to see that exact pairing again. And the facts that fade are the rare ones. Anything that comes up every few rounds is always fresh. A pairing that is rarely dealt is gone before it comes around again.

The other game

I'm autistic and I'm 58. Not knowing the rules is still how I go into most social interactions. Nobody hands you the scoring, so you work it out from what happens after you have played.

The reasoner needed 58 rounds. I've had 58 years, and I'm still short of the expert. I think the learners had it easier than I do, in three ways.

They get a screen. Every one of their rounds ends by telling the learner who won, with what, and for how many points. A conversation mostly doesn't. I've written before that I don't always know when someone is upset. In the game's terms, I don't always get the screen.

They get one rulebook. The game's scoring is the same in round 150 as in round 1. Mine isn't. When I think I've learned the rules, a new one is exposed, or the game changes. It can be a different kind of interaction, or the same kind with a different person, and part of what I learned doesn't carry over. The forgetful learners level off because facts leave as fast as they arrive. A rule that stopped applying costs the same as a rule I forgot.

And their game forgives them. This was the result I didn't expect. Whoever picks first has a large advantage, and the person playing always picks first. Between two experts, the first picker wins 99.5% of games. Even a careless first picker beats an expert two games in three. So a new player can know nothing about the scoring and still win. I suspect hiding the rules only works because of that. Nobody gave me first pick.

The simulation does not measure a person. The learners are a few hundred lines of code and the forgetting rates are guesses. But I recognize the curve: a learner who forgets, in a game with hidden scoring, levels off below the expert and never gets past that point no matter how long it plays.

No comments yet