Heat Alert Optimization#
$ cat projects/heat-alerts.md
Optimizing Heat Alert Issuance with RL
Cities issue heat alerts to protect vulnerable people. The problem: they issue them based on fixed temperature thresholds. That’s a blunt instrument when what actually matters is a complex interaction of weather, demographics, historical exposure, and alert fatigue.
This project reframes heat alert issuance as an RL problem. The environment is built from real data — 15 years of daily weather, Medicare hospitalization records, and county-level demographics across 2,500+ US counties. The reward model is Bayesian, because the causal signal is noisy and heterogeneous across locations.
The first surprise: unconstrained RL fails hard. It either never alerts or always alerts. The fix is constraint-based — budget the number of alerts per season and let the policy learn when they matter most. The constrained policy outperforms the National Weather Service’s current approach.
This is RL applied to real public health infrastructure, not a simulated benchmark. The kind of problem where getting the reward function right matters more than the algorithm. AAAI 2025.
$ cat /etc/motd
// …built with Claude Code. based on data, but verify.
$ cd ~