The Taunton projectTwenty AI ants, one street map, and a unanimous wrong answer
Build an ant colony in Python, let it choose a route from Taunton Library to Vivary Park, then check it against the exact shortest path.
The street map comes from the OpenStreetMap API: central Taunton turned into 1,044 junctions joined by 1,270 links. Taunton Library and Vivary Park are attached to their nearest junctions, which are 384 m apart in a straight line. An exact algorithm, Dijkstra's, examines 465 junctions and reports the shortest walking route: 523.4 m. That is the answer to beat.
Each virtual ant starts at the library and wanders from junction to junction, never revisiting one, choosing its next link at random but favouring short links and links with more scent on them. Ants that reach the park lay scent along their whole route, more for shorter routes, and after every round some scent evaporates so that old choices can fade. Twenty ants run for 100 rounds, and the experiment is repeated with 10 different random seeds. For the first rounds most ants get lost in dead ends; in the median run the first ant only reaches the park in round 21.
Where the ants finished after 100 rounds in each of 10 runs (Python, OpenStreetMap data, 29 September 2026)| Question | Answer |
|---|
| Exact shortest route (Dijkstra) | 523.4 m |
| Runs where the colony found that route | 0 of 10 |
| Runs where all 20 ants agreed on one route at the end | 10 of 10 |
| How much longer the agreed route was | 195.8 to 337.5 m (median 222.0 m) |
| Runs where a better route was found and then abandoned | 4 of 10 |
The colony always reached a decision. In every one of the ten runs, all 20 ants ended up walking exactly the same route, a perfect consensus. And in every run that route was wrong, typically 222.0 m longer than the shortest, about 42% further. Worse, in four runs the ants had actually found a shorter route earlier on, and then drifted away from it as the scent on their favourite path drowned it out. Tuning did not rescue them: across 24 more runs with more ants, faster or slower evaporation, or a stronger pull towards short links, the colony never found the 523.4 m route, and its closest was still 163.1 m too long.
A likely cause shows up in the logs. The first ants to arrive, after long wandering routes, laid the first scent, and every later ant was nudged along those same streets. The swarm amplified its first idea instead of searching for a better one, a trap known as premature convergence. Simple fixes exist, such as always remembering the shortest route ever found, but the deeper lesson is that agreement among agents is not evidence of being right.
Ages 8 to 11
Play a follow-the-trail game on a paper map and see how quickly everyone copies the first path.
Ages 11 to 15
Program a single random-walking ant in Python and count how often it reaches the park.
Ages 15 and up
Build the full colony, measure consensus and add a "remember the shortest" rule to fix it.
OpenStreetMap streets, our ants
Street data is from OpenStreetMap and its contributors under the Open Database Licence. The colony, the shortest-path check and all the route measurements are our own work, describing the mapped network inside one rectangle.
