Published optimisation research + real systems
I co-authored an IEEE paper (SCIS&ISIS 2018) on the Probabilistic Orienteering Problem: routing under uncertainty, where a stop might be gone by the time you reach it. We trained a neural network to predict how much Monte Carlo sampling a candidate route needs.
The Probabilistic Orienteering Problem is a routing problem with a catch. It looks like a travelling-salesman tour: visit a set of nodes, keep the route short. Except any node might turn out to be unavailable by the time you reach it. Picture a truck out delivering parcels, where some stops fall through on the day. You have to pick a tour that pays off on average, across every way the day might go.
Scoring a single candidate tour under that uncertainty means Monte Carlo sampling: simulate the tour many times over and average the outcome. More samples buy a more precise estimate and cost more time, and the right number of samples depends on the instance. Our contribution was to let a model make that call. A feed-forward neural network, trained as a regressor on features of each instance, predicts how many samples are enough to score a tour reliably. Around fifty samples brought the estimation error under 1%, with the standard deviation under 5%.
I wrote the implementation in C++ and tested it on the standard benchmark instances for the problem. The work was published as: Roberto Montemanni, Federico D’Ignazio, Xiaochen Chou, Luca Maria Gambardella, “Machine Learning and Monte Carlo Sampling for the Probabilistic Orienteering Problem,” SCIS&ISIS 2018 (IEEE), pp. 14–18, presented in Toyama, Japan. doi.org/10.1109/SCIS-ISIS.2018.00014.
It stayed my one published paper, and it settled a question for me: I’m happier building systems that run than doing research full-time, and I moved toward industry not long after. A couple of those systems came out of the same IDSIA period: the drone trajectory tracker and the rock-paper-scissors classroom classifier.