← Back to Sea Salt & Paper

About Sea Salt & Paper

Sea Salt & Paper is a quick card game of collecting sets, playing pairs and pushing your luck, with sea creatures folded from paper. It was designed by Bruno Cathala and Théo Rivière and is published by Bombyx. This is an unofficial fan-made version and isn't associated with the real game.

Like it? Go buy the real thing!

This is the two-player game with both expansions: Extra Salt's eight new cards and Extra Pepper's twelve events. The rules play the way they do on Board Game Arena in its Arena mode.

How to Play

This is the gist. For the full rules, read the official rulebook (PDF).

Objective: be the first to reach 40 points. A game takes a few rounds, and you score at the end of each. Placing all four mermaids wins the game on the spot.

Setup: all 66 cards are shuffled into a deck, and the top two are turned face up to start two discard piles. Your hand starts empty.

Your turn:

Card points count everything you have, in your hand and in front of you: every pair scores 1 (played or not), and collections, multipliers and mermaids score as their cards say (all of them are listed below). Most of the game is deciding which of these to build toward, and when to stop.

Ending the round: you can end it two ways.

An empty deck: if the deck has run out when a turn passes to the other player, the round ends and nobody scores.

The end: once someone has 40 points after a round, the most points wins. On a tie, whoever scored more in the last round wins.

The Cards

All 66 of them: the base game's 58, and Extra Salt's 8 (marked). Every card has a colour, which matters: the mermaids and the colour bonus count colours, and so does the Angelfish event.

Loading the cards…

The Events

Extra Pepper adds twelve event cards. At the start of each round the next one is turned up, and its rule applies to both players for that round. At the round's end it goes to the leader or to the player behind, as its arrow shows, and from then on its rule is theirs alone, for as long as they stay in that position. A player keeps one event at a time: receiving a second, they choose which to keep. On a tie, the event is discarded.

This Version

It's two players only, against the computer. A few things differ from playing at a table:

A Known Bug: Undo Lets You Cheat

On the vs AI page you can undo any decision, even after the computer has played. The order of the deck is fixed when each round is dealt, and undoing doesn't reshuffle it. So you can draw two cards, see them, undo, and then decide whether to draw or take from a pile, knowing what's coming.

I don't plan on fixing it. Whether you cheat this way is up to you. The computer never does: it only knows what a player in its seat would know.

The AI

It works like DeepMind's AlphaZero. A neural network learned the game by playing against itself, with no human games or tips. It started out knowing only the rules and playing at random. Before each decision, a tree search looks ahead 100 to 400 times, using the network to pick out promising lines and to judge how things stand where they end. The network then learns from the decisions the search made and from how the games turned out, and the better network plays the next games.

Hidden cards: nobody at the table can see the other player's hand or the order of the deck, and the computer can't either. It's given only what a player in its seat would know: its own hand, the table, every card that went onto the discard piles, and the cards it saw go into your hand. Each time its search looks ahead, it first deals out the cards it can't see in a way that fits what it knows, and plays the line out in that deal. After many guesses, it has a sense of what usually happens. Besides who will win the game, the network also learns to predict how the current round will be scored.

What it learned: by iteration 830 of its training it had played about 1.8 million games against itself, in about two days on one graphics card, and it won about four games in five against a search that plays without the network. Its style changed along the way. At first, half its rounds ended with STOP. Now it says LAST CHANCE in about 60% of rounds, and wins about three quarters of those bets. Wins by placing the mermaids went from 2% of its games to about 12%, and its games got shorter, from almost six rounds to about four.

It all runs in your browser. Medium searches 100 times per decision, Hard 400 (as in training), and Expert as many times as your device can in 1.5 seconds, so it's a bit stronger on a faster device. Easy doesn't use the network: it takes whatever adds the most points right now, plays every pair, and says STOP as soon as it can. Random plays any legal move. The blue and orange bar during the game is the search's estimate of each side's chances. It only moves when the computer thinks, so it never reacts to your own moves.

The Analysis page follows a game from your seat, such as one you're playing on Board Game Arena. You enter your decisions and what you see the opponent do, and it analyzes each of your decisions from what you know. It never sees the opponent's hand unless you do.