About Radlands

Radlands is a two-player card game set in a post-apocalyptic wasteland. Each player defends three camps with people and events, paid for with water. It was designed by Daniel Piechnick and is published by Roxley. This is an unofficial fan-made version and isn't associated with the real game.

Like it? Go buy the real thing!

How to Play

This is only the gist of the rules, and it leaves a lot out. For the full rules, read the official rulebook (PDF). Every card also explains itself: point at one in the game to read it.

Objective: destroy all three of your opponent's camps before they destroy yours. If the draw deck runs out a second time, the game is a draw.

Setup: you're dealt six camps and keep three, one at the back of each of your three columns. Each camp has a number, and they add up to the size of your opening hand. This version deals the camps of the base game and of the Cult of Chrome expansion, all shuffled together.

Your turn:

• The event at the front of your queue goes off, and the rest move one space closer.
• You draw a card and get 3 water. Water you don't spend is lost at the end of your turn.
• Then spend your water however you like: play people in front of your camps (two to a column), play events into your queue to go off on a later turn, or use the abilities of your people and camps. You can also junk a card from your hand for the junk effect printed on it, or pay 2 water to draw a card.

Damage: most attacks can only hit an unprotected card, the front one in a column, so your people stand guard over the camps behind them. A damaged card is destroyed if it's damaged again, and a damaged person can't use its abilities until it's restored.

Raiders: every player has a Raiders event. Raiding sends it into your queue, and when it goes off, your opponent has to damage one of their own camps.

A Known Bug: Undo Lets You Cheat

On the vs AI page you can undo any move, and that makes it possible to cheat against the computer. Undo works even after the computer has taken its turn, and even on a move that drew cards. The order of the deck is fixed when the game starts, and undoing doesn't reshuffle it.

So you can see what's coming. Draw a card (for 2 water, or with any ability that draws), look at it, undo, and play knowing what you'll get. Or end your turn, watch what the computer does and what you draw next, then undo and replay your turn knowing all of that.

This is a known bug, and I don't really plan on fixing it. Whether you cheat this way is up to you. The computer never does: it only knows what a player at the table 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, and trained in a loop:

• The network looks at a position from one player's seat: their hand, both boards, the discard pile, and how many cards are hidden where. It gives two answers: which moves look good, and how likely it is to win, draw or lose.
• Self-play: the network plays itself, thousands of games at once. Before every decision, a Monte Carlo tree search looks ahead 100 to 400 times, using the network to pick out promising lines and to judge where they end. The search plays better than the network on its own.
• Hidden cards: nobody at the table can see the opponent's hand, the order of the deck or what a face-down punk is. So before each look ahead, the search makes a guess: it shuffles every card it can't see into a deal that fits what it knows, and plays it out there. Many guesses later, it has a sense of what usually happens.
• Learning: after every 100,000 new positions, the network trains on the most recent 1.5 million. It learns to suggest the moves its search chose, and to predict how the games turned out. The better network then plays the next games, its search gets better too, and the loop goes on.

The first try taught a lesson. A Radlands position nearly gives away which game it came from, so a network that learned the result from every position just memorized which games were won, instead of learning to judge positions. The run playing here learns the result from only four positions per game, and the search's own estimate at every position. By the time of this network it had played over a million games, in a few days on one home graphics card.

It all runs in your browser. Medium searches 100 times per decision, Hard 400, and Expert as many 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 guesses the hidden cards once, tries every move, and takes the one that leaves it the most on the table. Random plays any legal move. The blue and orange bar during the game is the search's estimate of each side's chances, counting a draw as half a win.