Catan · AlphaZero Play vs AI

Superhuman Catan AI

We took the AlphaZero algorithm that beat Lee Sedol in Go, trained it on Catan from zero human games, and recorded 15,000 full matches.

290,000 Training Games 15,000 Library Games 12.1M Logged Events
Play Against the AI
Runs in your browser via WebAssembly & ONNX · No install
Scroll to explore

The Algorithm

Seoul, 2016.

In 2016, AlphaGo beat world Go champion Lee Sedol 4–1. Older programs searched by brute force. Go has 250 legal moves per turn, so brute force failed. AlphaGo joined neural intuition with tree search instead.

Self-Play

Zero Human Games.

One year later, AlphaZero dropped human games completely. It started from random weights and played itself. Tree search found better moves. The network trained to match that search. Repeat the loop, and skill appears.

The Island

Dice. Hidden Cards. Four Seats.

We built the exact same loop for Catan. Our model is a 79-token Transformer over the hex graph. Before each search, it samples a valid hidden deal from public counts. It never peeks at opponent hands.

Training

Ten Hours. From Noise to Mastery.

On one GPU, the AI played 290,000 games in 9.6 hours. Even with search turned off, its win rate against three rule-based bots climbed from 0% to over 50%. Even parity in a four-player game is 25%.

Setup

First Seat Leads.

Next, four copies of this AI played 15,000 full games. All four seats had equal skill. Seat 1 won 26.8%. Seat 4 won 23.9%. Going first is worth about one extra win every 35 games.

Setup

Cover All Five Resources.

Resource count is the strongest opening signal in all 60,000 starts. Openings on all five resources won 33.5%. Openings on three won 17.8%. Openings on two won only 10.0%.

Setup

Never Skip Wheat.

Missing wheat hurts most. Openings without wheat won 19.2%, versus 26.0% with wheat. Brick is skipped most often (30.7% of starts), and missing it costs almost five points.

Setup

Variety Beats Raw Pips.

Hold pips fixed and compare resource count. A thin start under 18 pips on five resources won 19.7%. That beat a rich 21–23 pip start on three resources (18.2%). Spread first, chase big numbers second.

Setup

Pips Flatten at 22.

Pips still matter up to a clear ceiling. Win rates climb fast from 15 pips (9.4%) to 22 pips (28.7%). Past 22 pips, extra starting production stops adding wins.

Early Game

Build to a Harbor. Do Not Start on One.

Starting on a harbor won only 22.9% because coast spots touch fewer tiles. Building to a harbor in turns 1–19 won 41.9%. Start inland for pips, then expand to a port.

Early Game

Production Wins Two in Three.

Rank the four players by total cards rolled by the dice. The top producer won 64.3% of games. The lowest producer won 3.0%. Winners averaged 72 dice cards versus 55 for others.

Early Game

Upgrade Early.

Players who built a city by turn 9 won 40.6%. Players who never built a city won 4.7%. Winners reached 5 points at turn 36 on average — nine turns ahead of the table.

Dev Cards

Monopoly Is the Jackpot.

We checked the AI's own win estimate before and after 314,000 draws. Monopoly raised win odds by +1.03 points above an average draw. Every other card moved win odds by 0.12 points or less.

Dev Cards

Name Ore or Wheat.

When the AI played Monopoly, it named ore (7,606 times) and wheat (7,147 times) most often. Buying more dev cards also tracked wins: players who bought 10 or more won 42.8%.

Trading

Nobody Helps the Leader.

Across 2.6 million trade answers, players accepted only 17.3% of offers from the leader. They accepted 28.3% from a player 3+ points behind. Nobody coded this rule. Self-play invented it.

The Robber

Block the Leader's Five.

The robber targeted the leader on 48.0% of all steals. The AI blocked 5 tiles (13.6%) as often as 6s (13.6%) and 8s (13.4%). A solo 5 hurts the leader without blocking yourself.

Endgame

Cities Explain Half the Gap.

Winners ended on 10.05 points. Other players ended on 6.10. Cities supplied +2.03 of that 3.95-point gap. Longest Road won 62.4%. The top winning board was 3 cities and 1 settlement.

Endgame

Leads Hold Early.

At turn 24, the eventual winner was already in 1st place 59% of the time, and in 4th place only 3.7%. At turn 40, a 3-point lead won 68.5%, and a 4-point lead won 85.0%.

Interactive Lab

Test Your Opening.

Pick your starting resources, total pips, seat, and harbor. The gauge shows how similar openings scored across 60,000 starts.

Resources touched4 / 5
Total opening pips20 pips
Seat
Start location
26.4%
+1.4 pp vs 25% parity

Five resources.
Early cities.
Never feed the leader.

No strategy rules were written into this AI. Every lesson above emerged from 290,000 games of self-play.

Play Against the AI
Challenge three copies of this exact model in your browser