The idea
The universal swarm study found the same three ingredients in real swarms: align with neighbours, switch from push to pull at about one neighbour distance, and pull the group back toward a set point. Each framework below reuses those rules as a controller, and each is compared with the simple rule or forecast people use today.
Results
| Domain | What the framework does | Data | Result |
|---|---|---|---|
| Energy grid | batteries share discharge by state of charge (alignment) toward a 24-hour set point | real ERCOT load, Oct to Dec 2020 | the rule-based framework cuts demand swings by 7.3% but not the peak. On the same data the trained battery swarm (10 seeds) cuts swings by 10.8% and the peak by 704 MW, ahead of the proportional rule (9.1%), a 24-hour forecast planner (6.8%) and a threshold rule (5.0%); the perfect-foresight bound is 949 MW. Use the trained swarm |
| Mixed-role energy | batteries, solar with curtailment and demand response each act on the net load they see, with no coordinator | real ERCOT + NREL data until the Ontario and Alberta files are downloaded | demand swings −14.5% and peak −453 MW with all three roles, against −11.7% with solar and batteries. More |
| Wildfire | each fire forecasts its final size with a growth law and bids for crews | real US and Canadian fires (203 fires, the same fires at every horizon) | median log error after 5, 8, 12 and 16 days: growth law 1.81, 1.52, 0.94, 0.68; persistence 0.83, 0.84, 0.56, 0.40. More days help the growth law a lot, but simple persistence still wins; it needs weather and fuel inputs |
| Epidemics | each region forecasts its wave and asks for help by forecast growth | real COVID-19 first wave, forecast 21 days ahead from the first 20% of the wave | Gompertz 0.16, Richards 0.26, exponential 0.86 (median absolute log error): saturating laws are 3 to 5 times better |
| Air traffic | aircraft add swarm separation and alignment only inside an alert radius | real OpenSky traffic for density; simulation for control | 0.51 close pairs (under 5 NM) per real snapshot; in simulation, close-pair steps fall from 28,635 to 3,920 (−86%) |
| Drones | flocking controller with gains measured in real swarms, plus goal seeking | simulation | all 30 drones arrive; 3 obstacle hits vs 18 without alignment |
| Logistics | ant-style task pickup with thresholds that adapt | simulation | mean queue 5.6 vs 69 for round robin, with 70% fewer moves |
What this does and doesn’t show
- The epidemic and air-traffic results are the strongest: real data, a clear baseline and a large gap.
- The drone and logistics results are simulations. They show the rules work as controllers, not that they beat a tuned real-world system.
- The wildfire forecast is an honest loss, and the rule-based grid framework is weaker than the trained battery swarm. They are listed so the next step is clear.
What's next
- Wildfire: add weather and fuel inputs, since the growth law alone still loses to persistence.
- Energy: rerun the mixed-role swarm on Ontario (IESO) and Alberta (AESO) data.
- Rule first: a swarm controller stays on only if it beats the rule on held-out data.