Skip to main content
  • Work
  • Experience
  • About
Get in touch
  1. Work/
  2. SomaHouse
Research2026

SomaHouse

A small circuit taken from the fruit fly connectome learns to run a simulated house, then goes up against conventional controllers. The wins and the losses are both in the report.

Role
Independent research
Outcome
Two registered simulation studies completed
Link
Code, report and trained models on GitHub

Project notes

SomaHouse began with a simple question: could a small circuit taken from the fruit fly connectome learn to look after a house? A house is a good test because its jobs depend on each other. Opening a vent changes the air and the heating load, and sunlight through a window changes both the room temperature and how much lighting is needed. I wanted to see whether a brain built from separate parts that pass messages to each other could handle that.

I built a simulated three-room house with changing weather, daylight, occupancy, vents, lights, and limited heating and cooling. The controller is split into four sections for temperature, air, light and context, and they share information through a small central hub. The wiring inside each section comes from real synapse counts in the MaleCNS fly connectome. The sensors, the hub and everything that connects to the house are my own engineering choices. I trained 24 versions across three random seeds, froze my pick of model, and only then ran the final tests.

The results were modest. Doubling the circuit gave 4.1% lower cost than giving the smaller one the same extra training, and cutting it in half made it worse. Removing the hub hurt too, so the sections really were using it. But a conventional balance controller still used less energy and met more of its comfort targets, and a dense network of the same size beat the fly-wired one. I ran a second study where the circuit first learns to copy a tuned conventional controller and then trains on reward. It finished 1.34% behind that controller, and my registered test for beating it did not pass. None of this shows that biological wiring is a better way to run a home.

What I enjoyed most was making it inspectable. You can follow a signal through the hub, trace every copied cell back to its source neuron, and watch the saved model run a 48-hour test in the film above. It is still a simulation, not a calibrated building, and nothing here is connected to a real home. It was a fun experiment, and it left me with a clear list of what to try next: a more detailed building model, hard limits on comfort and air quality, and better handling of faulty sensors.

Simulation study. Code is MIT licensed; connectome-derived data and trained checkpoints are CC BY 4.0.

Gallery

Browse

All work

Next

Passerby

P
Passerby street photography

© 2026 Michael Wong

Made with care in New York.

AboutGitHubLinkedIn