Machine learning experiment
HabFly
HabFly is a learning experiment built around a map of connections between fly neurons. I’m exploring how it can learn sequences of decisions in astronomy tasks, from choosing a calculation to reusing a result and entering an answer.
- My part
- I developed the learning engine, task environments, browser adapter, and Rust terminal interface.
- Research in progress
- Testing is ongoing. A longer article about the design, experiments, and results is coming soon.
HabFly action loop
- ObserveVisible task state
- DecideRecurrent graph
- ActTool / browser
- CheckRead back the result
What I’m exploring
The learning engine uses Python and PyTorch, with a recurrent graph whose wiring is based on connections between fly neurons. The policy chooses actions, while astronomy tools handle calculations and return results that can be reused in later steps.
A browser adapter connects those decisions to an astronomy activity. A Rust terminal interface lets me inspect decisions and network activity, step through a run, and replay recorded events.
Where it stands
I’m testing the learning system in local task environments and bounded browser trials. Full learned completion of the real HabWorlds project has not been established.
The forthcoming article will cover the architecture, what the experiments show, and the limitations that remain.