Swarm Intelligence has its own section at /swarm/, which I update as the work moves. This page is the short version for the projects list.
The problem
Many decisions happen at once: when each battery charges, where each drone flies, which province lends a fire crew. A central planner gets slow and fragile as the number of agents grows.
What I built
One engine in which every agent decides from what it can see and the group still acts as one team. Each application is compared with the rule people use today and, where possible, the best possible result.
Results
On a stressed grid, the swarm cut the peak by 704 MW on average over 10 training seeds (the perfect-foresight best is 949 MW) and reduced demand swings by 10.8%, against 9.1% for the best rule. On the public CityLearn 2022 benchmark it scored 0.96 against 1.03 for CityLearn’s tuned rule. In wildfires the simple rule still wins, and the site says so.
What’s next
The full write-up is the paper draft (PDF).
Natural swarms
The work now also studies real swarms from measured data only. Fish schools came first. Then the same measurements on 25 real datasets found one general law: agents align with neighbours, switch from keeping apart to drawing together at about one neighbour distance, and the group pulls itself back toward a set point (the study). Six application frameworks reuse it, and a speculative model asks whether the universe’s expansion is regulated the same way.
What’s next
Three tracks, all on /swarm/: applications in energy, oil and gas and wildfires; swarm agents working with AI agent crews; and the natural swarm study, next with large bird flocks.
Limitations
- All results so far are from simulations: real grid data from 2020 for the grid batteries, and simulated forests and hotels.
- In wildfires, no learned swarm has beaten the nearest-fire rule yet (87.2% against 90.1% of forest saved), and learned drones lose ground as the swarm grows.
What's next
- Energy swarm with batteries, solar and demand response on Ontario and Alberta data.
- AI agent crews that watch the swarm's runs and propose fixes a person approves.
Updates
No logged updates yet. New work on this project will appear here and in the activity log.