CATAN 1v1 BOT
About · How It Works
How It Works
Graph Neural Network (~1.25M parameters)
Rather than flattening the board into a list of numbers, the bot models it as a graph. Each hex, edge and settle spot exists as interconnected nodes, structured exactly as they are on the physical board. The GNN processes this structure through message passing, each piece "talks" to its neighbors in multiple rounds until every tile, corner, and edge has absorbed context from across the whole board. The network uses this to score candidate moves and estimate who is currently winning.
MCTS (Monte Carlo Tree Search)
The bot uses MCTS to think ahead, exploring hundreds of candidate moves before committing to the one that consistently leads to good outcomes.
Training
The bot was initialized by watching 45,000 rule-based bot games (~13 million unique positions), learning to predict the winner and imitate the moves played. It then competed against those bots until it surpassed them entirely. Today it trains exclusively through self-play and has played roughly 15,000 games against itself.
Quirks & Caveats
The bot sees your hand
The bot has perfect information it can see your resource cards and playable development cards at all times. This was how it was trained as MCTS in general preforms better with perfect information, in future generations I will try to fix this. In the mean time you can 'cheat' back and click the Bot Resources or Bot Dev Cards buttons in the game sidebar to see its hand too.
Balanced dice awareness
Colonist.io uses a weighted dice distribution to push numbers towards rolling closer to their expected frequency rather than being purely random. The bot's architecture includes next-roll probability and cards-until-reshuffle as explicit input features, so it can reason about when numbers are "due" in the same way top 1v1 players do.
Replay Guide
You can load any Colonist.io game into the analyzer to see the bot's policy and value estimates at every move.
Getting a replay link
To analyze a replay first copy the shareable replay link. This is pay walled on Colonist.io so you will either need a subscription or the game needs to be generated via the Catan Champs Discord bot
Step 1
Copy the colonist.io/replay/… URL from the post-game share button or the Catan Champs Discord
Step 2
Go to Analyze Replay from the home screen, paste the link, and click Load
Step 3
Click Analyze to run the bot over the full game, this takes ~30 seconds
Step 4
Once complete, every position has an eval bar, and you can show the top 3 candidate moves the bot would consider from any position
Step 5
Use the navigation controls to step through the game move by move, or jump to any position and resume play from that point to try different choices at key decision points