LastMileAdapt Re-optimization Policy Selector
Lightweight policy selector that recommends a re-optimization strategy during intraday last-mile delivery operations.
Purpose
Given fleet state features at a disruption point, predict the best strategy among:
| Strategy | When to use |
|---|---|
| Rolling Horizon | Active disruptions, need low route churn |
| Reinforcement Learning | High re-optimization frequency |
| Monte Carlo / Scenario Tree | High demand/traffic uncertainty |
| RouteFinder | Large instance, need fast full replan |
| OR-Tools | Tight time windows, feasibility focus |
| ALNS | High capacity utilization |
Input Features (12)
n_active_orders, n_vehicles_available, capacity_utilization, traffic_index, disruption_severity, time_progress, pending_injections, failed_vehicles, avg_time_window_slack, route_fragmentation, stochastic_demand_var, reopt_frequency
Files
| File | Description |
|---|---|
config.json |
Model configuration and objective weights |
policy_weights.json |
Linear policy weights and biases |
Objective Context
The selector optimizes for composite cost and schedule stability:
objective = distance + 8 × lateness + 12 × route_change_penalty
References
- Base routing model: ai4co/routefinder
- Benchmark dataset: MBZUAI/svrp-bench
- Simulator: adaptive-last-mile-routing
License
Apache-2.0