I build planning tools and test them on real operations.

Vaccines and blood products are collected at local depots in 40 cities and must reach any other city fast. Four hubs are to be opened, chosen from 12 candidate cities.

A p-center MIP picks the hubs that minimize the longest depot-to-hub trip. A small squared-distance term breaks ties toward compact clusters, which makes the routing step easier.

Each hub’s routes are solved exactly where the MIP is fast enough, and with Clarke-Wright savings plus 2-opt where it is not, under a 12-hour shift limit. Then location and routing are solved together under an 8-hour limit.

The sequential design needs 7 trucks (hubs in Adana, Ankara, Kocaeli and Muğla). Solving location and routing together cuts the longest delivery from 27.1 to 19.4 hours, but it needs 15 trucks.

UMKE, Türkiye’s national medical rescue team, needs field health facilities placed so that all 20 affected districts are within a medical service radius.

Three views of one decision: set covering (fewest sites that reach everyone), maximum coverage (most people reached with at most 4 sites) and p-median with a coverage limit (shortest total distance).

Each model is solved for service radii from 2 to 6 km, to show the trade-off between the number of sites and the distance people travel.

Full coverage needs 5 sites at 2 km but only 3 at 3 km. With 4 sites, coverage is 92.9% at 2 km and 100% from 3 km on, and the same four districts stay optimal from 3 to 6 km.

Late deliveries often come from drivers waiting in the wrong place. With 500 drivers, how many should wait in each of Manhattan’s 32 neighbourhoods, and how far should each one serve?

A binary LP chooses one of five coverage radii for each neighbourhood, with effective demand growing with the square of coverage. Drivers left over after the minimum plan are spread in proportion to hourly demand.

VBA modules compute neighbourhood centroids with the shoelace formula and Manhattan distances to restaurants, and cache them so the model does not rebuild its inputs on every run.

Caching cut model setup from minutes to seconds. An Excel form takes the number of drivers and shows the allocation on an interactive Leaflet map.

In FetchReach-v4 a simulated arm gets a reward of −1 at every step until its hand reaches the target. Almost every random attempt fails, so there is very little to learn from.

Soft Actor-Critic with Hindsight Experience Replay, which relabels a failed attempt with the goal the arm actually reached, so every attempt teaches something.

The course provided the critics, the replay buffer and the training loop. I implemented the actor, a Gaussian policy with the reparameterization trick and the tanh log-probability correction, trained it for 10,000 episodes and ran three one-change ablations.

The trained agent reached the target in all 100 evaluation runs (average reward −2.39). A 10× lower learning rate stopped learning entirely, and a higher entropy weight left the arm moving randomly near the target.