sim2find

Turning clues into search maps for missing people with dementia.

Comment

Inspiration

We wanted to work on missing person cases that go unsolved because the evidence is confusing or there’s almost none. That was too big for 36 hours. But while reading, one group kept showing up: older people with dementia who walk away from home and get lost. They don’t know they’re lost, so they don’t call for help or answer when people shout their name. They just keep walking, and every hour they’re out there, the chance of finding them alive goes down. That stayed with us.

What surprised us is how much is already known about them. Search-and-rescue teams have collected data on more than a thousand of these cases: how far people usually go, where they’re found, and that most of them just keep walking in one direction until something stops them. But when someone actually goes missing, the search often still starts with a circle around the house. We wanted to see if we could turn that knowledge, plus whatever clues come in, into a map that tells a team where to look first.

What it does

You tell Sim2Find where and when the person was last seen. It runs 20,000 simulated walks from that spot on the real Amherst map, based on how people with dementia usually behave. The result is a heat map, not a pin: red areas are where most walks ended, so that’s where to look first.

When a clue comes in, the map updates. A sighting pulls the heat toward that place. A camera that saw nothing, or an area already searched, cools those spots down. You can type a witness report in plain words and an LLM turns it into a clue, but a person always checks it first.

How we built it

We built a 10 m grid of Amherst from OpenStreetMap, elevation data, and land cover, so the walkers know about roads, buildings, fences, woods, swamps, and streams. Each walker follows rules from Robert Koester’s lost-person research on dementia cases. Most go straight until they get stuck, many cross roads instead of following them, and they tend to go downhill.

The hard part was making the rules honest. We didn’t want to just pick numbers that looked good, so we used Optuna to tune them until our walkers matched 16 real statistics from that research: how far people go, and where they end up. Then we tested it on 999 cases where the true location came only from those statistics, not from our own walkers. With one sighting, Sim2Find beat circle search 83% of the time and needed 56% less ground.

Challenges we ran into

We built a 10 m grid of Amherst from OpenStreetMap, elevation data, and land cover, so the walkers know about roads, buildings, fences, woods, swamps, and streams. Each walker follows rules from Robert Koester’s lost-person research on dementia cases. Most go straight until they get stuck, many cross roads instead of following them, and they tend to go downhill.

The hard part was making the rules honest. We didn’t want to just pick numbers that looked good, so we used Optuna to tune them until our walkers matched 16 real statistics from that research: how far people go, and where they end up. Then we tested it on 999 cases where the true location came only from those statistics, not from our own walkers. With one sighting, Sim2Find beat circle search 83% of the time and needed 56% less ground.

Accomplishments that we're proud of

A lot of things went wrong. Our walkers kept stopping too close to home. We were sure it was buildings blocking them, but when we actually measured it, the real problem was how we modeled walking time. We also couldn’t get people to end up on roads: our model said 5%, the data said 18%. It took us a while to realize the data counts people who were spotted by drivers while still walking. So we added that, and in the app, “nobody has reported him yet” became a clue.

Our first test was also humbling. For people found more than 3 km away, our model lost badly. We fixed it by starting new walkers from the sighting point instead of only reweighting the old ones. That one change turned most of those losses into wins.

What we learned

Measure before you fix. Every time we guessed why something was broken, we were at least partly wrong. We also learned to label every rule as data, expert behavior, or our own assumption, because the first question anyone asks is “where did that number come from?”

What's next for sim2find

The test cases are built from published statistics, not real records, and the demo case is made up. The next step is working with a real search team and checking it against cases they’ve actually solved.

Built With

Share this project:

Updates

Submission history