A generals.io FFA Bot

Overview

generals.io FFA is a real-time strategy game: each player starts with one general on a grid, grows an army, and spreads across land and cities, around impassable mountains, under fog of war. Capture a rival’s general and you absorb everything they own. I wanted to build a bot for it, but the first move wasn’t the bot. It was the world it will learn in.

What I built

It starts with the simulator, an engine matched to generals.io’s rules for 2-8 players: army growth (generals and cities tick up every turn, on every 25th turn every owned tile grows), movement and capture, neutral cities, and per-player fog of war that reveals only a 1 tile around own general tiles. Map generation places generals with guaranteed mutual reachability, and the whole game is deterministic from a seed.

On top sits a heuristic commander, a priority ladder: defend the general first, then grab adjacent land, take an affordable neutral city, otherwise gather the largest army and march it toward the nearest enemy, remembering where enemy generals were last seen even after they slip back into fog. Pathfinding is plain breadth-first search over passable tiles.

To know whether a bot is any good, I built an evaluation system, it plays seeded games, ranks finishers by survival, then land, then army, and converts those placements into pairwise Elo updates on a bot ladder.

The learning layer: a pipeline that trains a network to imitate the heuristic first afterwards learning with self-play. Each position is encoded, moves into an eight-directions-per-tile action space with a legality mask, and a fully-convolutional policy-value network (size-agnostic, so it plays any board) learns the heuristic’s moves against a placement-based value target.