Hackathon · Oct 2026
Hyde AI Builder Hackathon
A Catan-playing agent that fits in one CPU core and 1.5 seconds per game.
- Result67% win rate vs the reference bot
- WhereIIT Bombay, team catana
- Budget1 CPU core, 2 GB, 1.5 s per game
Problem
Build an agent for two-player Catan on the organisers' engine. Every agent gets one CPU core, 2 GB of memory and a clock of 300 ms plus 12.5 ms per decision, capped at 1.5 seconds for the whole game.
Agents are scored on strength against two reference bots and on robustness against the weaker half of the other teams.
What we built
- A whole-turn beam search (trade, build, buy) that runs inside the clock and never peeks at dice or hidden cards.
- A tiered evaluation of the position: production, hidden victory points, races for Longest Road and Largest Army, ports and discard risk.
- Two extra terms: denial, the resources the opponent cannot produce, and concentration, how much of our income one robber can block.
- A small ensemble of MLPs that only breaks near-ties between the top two moves.
- An arena that matches the official scoring, and a pool of opponents to tune against.
Design choices
- Shallow search with a well-shaped evaluation, because the clock rules out deep search and large networks.
- Tuned against a pool of opponents rather than one bot, since robustness is scored against other teams.
- A change was kept only if it beat the previous version outside its 95% interval on unseen boards.
Result
67.2% win rate against the reference bot over 2,000 unseen boards (95% interval 65.2–69.3%).
On the organisers' leaderboard: no games over the clock and no faults.