The problem. A fire spreads unpredictably with the wind. A small fleet with limited payload has to decide, again and again, which flank to attack, when to go back and refill, and how to avoid all crowding onto the same spot.
The swarm. Each drone carries 6 loads of retardant and sees a 7×7 patch of forest around itself. Every step it moves, drops retardant or holds, and it refills on the launch pad in the top-left corner. Two fires start in random places with a random wind on a 16×16 forest. The drones move three cells for every step the fire spreads. The team is rewarded for every cell of forest that doesn't burn.
Watch a fire
Recorded runs on unseen fires. Every approach faces the same three fires. Press play, or drag the slider.
Cells on fire at each step, this fire
All approaches compared
Averages per fire over 50 unseen fires.
| Approach | Forest saved | Cells burned | Most cells on fire at once | Steps to put out |
|---|---|---|---|---|
| Nearest-fire ruleeach drone flies to the closest fire | 90.1% | 25 | 10 | 32 |
| Swarm | 87.2% | 33 | 11 | 39 |
| Random moves | 36.8% | 162 | 35 | 151 |
| No drones | 33.2% | 171 | 38 | 150 |
How the drones were trained
Forest saved on the same 50 fires, by training approach.
Honest result. The best drone swarm saves 87% of the forest. The simple "fly to the nearest fire" rule saves 90%. Learning from scratch only reached 77%. Letting the swarm first copy the rule and then improve with practice closed most of the gap, but it hasn't overtaken the rule yet.
Longer training and drone-to-drone messages did not help in this setup (85% and 80%). Those are useful negative results that point the next experiments elsewhere.
What comes next
- Real fuel and terrain maps (Canadian Wildland Fire Information System) instead of a uniform forest.
- Several launch pads. With one pad, results fall as the forest grows to 100 and 1,000 drones, and the simple rule falls too.
- Situations where a fixed rule breaks: crossing fires, wind shifts, drones with different payloads.