Generating practical last-mile delivery routes using a data-informed insertion heuristic
Bibliographic record
Abstract
Couriers often deviate from pre-planned delivery routes due to practical realities that routing algorithms may overlook. We statistically demonstrate that, in addition to travel time, factors like turn sharpness, backtracking distance, and neighbourhood visit timing influence a driver’s navigational choices and propose a Data-informed Insertion Heuristic (DIIH) for Travelling Salesman Problems (TSPs), which considers a custom cost function inferred from historical routes. The DIIH is trained on the dataset from Amazon’s Last-mile Research Challenge, which contains historical TSP instances classified into one of three qualities: high, medium, or low, indicating the satisfaction level of Amazon’s logistics planners of a route based on productivity, courier experience, and customer satisfaction levels. We train an energy-based model to predict the likelihood of a route being of high quality. Compared to existing benchmarks, the DIIH generates 22.4% and 24.1% additional high-quality solutions than the Amazon challenge winner and courier-performed routes, respectively. This improvement comes with an increase of 20.6% and 13.9% in the median travel time. While optimizing purely for travel time would result in shorter routes, we account for both travel time and human preferences, which explains the observed tradeoff. We show that the probabilistic evaluation of a route measured by the energy-based model developed in this study is a promising metric for estimating a routing algorithm’s practical performance. • Learns driver preferences to improve last-mile delivery routing. • Introduces a heuristic using backtracks, turn angles, and visit timing. • Predicts route quality using a logistic regression classifier. • Outperforms benchmarks with 22%–24% more high-quality routes. • Framework extends to dynamic and pickup-delivery TSP variants.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".