Competition · Kaggle
When a learned policy beat the hand-tuned one
Orbit Wars is a multi-agent reinforcement-learning competition. I entered with a learned, emergent-control policy that reached a best score of 406.6.
The more interesting result wasn't the number — it was that my hand-tuned strategies kept underperforming the learned ones. That became a design principle I carried straight into Aura: keep the presets minimal and let the behaviour emerge. Orbit Wars is where "Zero Presets" stopped being an opinion and became something I'd measured.