P4Learned Delivery Routing

Intelligent Systems · Project P4 · November 2026

Learned delivery routing

A neural network learns how long trips take across New York at every hour. A constraint solver gives each order a van, and A* orders each van's stops. When a route runs late, the system learns a new rule and plans again.

Moaz Hassan El-Garawany · Ahmed Hossam El-din

SlidesReportSource
Three van routes across Manhattan and JFK on the demo day
Demo day: three vans, twenty orders, every one on time.

The dossier

Thirteen items on five lines. Each stop opens the work itself.

Code and data

  1. 1CodebaseOne package split by responsibility, four entry points, a pytest suite
  2. 2Data pipelineNYC taxi trips, January 2022, filtered and sampled with a fixed seed
  3. 3Data dictionaryEvery column, real or generated, and how it was made

Models

  1. 4Travel-time modelA numpy MLP measured against three baselines
  2. 5Route searchUCS, Greedy and A* with node counters and a brute-force check
  3. 6Van assignmentA CSP with switchable backtracking, forward checking and AC-3

System

  1. 7Integrated demoOne command: input, solution, trace, pruning, model on and off
  2. 8SetupPinned environment and the exact commands, in order
  3. 9Test suite29 test functions, 38 cases, normal, edge and impossible

Evaluation

  1. 10ExperimentsThree experiments, ten seeds each, mean and standard deviation
  2. 11ReportTen pages: problem, design, results, limits, conclusions
  3. 12PresentationTen slides with speaker notes and the live demo script

Team

  1. 13ContributionsIndividual statements and the git log