Solving the Takeout Delivery Problem: Balancing Efficiency with Human-Centric Logistics

TL;DR

As the O2O (Online-to-Offline) food delivery market saturates, the competition has shifted from user acquisition to operational excellence. This paper proposes a novel framework for the Takeout Delivery Problem (TDP) that optimizes for both platform costs and customer satisfaction while introducing a Human-Computer Interaction (HCI) loop that allows riders to provide feedback on system-generated routes.

Motivation: The Hidden Cost of "Beautiful Data"

While delivery giants like Meituan and Eleme report impressive statistics—shorter delivery times and higher rider incomes—there is a darker side to the "intelligent scheduling" craze. Standard algorithms often ignore real-world friction, leading to:

• Rider Burnout: Rigid systems treat humans like robots, causing high turnover.
• Social Friction: Unrealistic deadlines forced by the system lead to traffic violations and safety risks.
• Rigidity: Purely mathematical models cannot account for the "ground-truth" experience of a veteran rider.

The authors argue that the TDP is more complex than a standard VRPPDTW (Vehicle Routing Problem with Time Windows and Simultaneous Pickup and Delivery) because rider starting locations are not fixed, and customer satisfaction involves both time and food quality (e.g., temperature).

Methodology: The Two-Stage HCI Strategy

The core innovation lies in a bi-objective model solved in two distinct stages.

1. The Bi-Objective Model

The model minimizes a function combining:

• Operation Costs ($Z_1$): Depreciation, fuel, and penalty costs for time-window violations.
• Satisfaction ($Z_2$): A fuzzy evaluation of delivery time and food quality.

$\min f=a(Z_{11}+Z_{12}+Z_{13})-b(w_1{Z}_{21}+w_2{Z}_{22})$

2. Algorithmic Architecture

To solve this, the authors utilize a Two-Layer Coding Method (linking specific orders to specific riders) and an Adaptive Multi-population Genetic Algorithm (AMGA).

• Stage 1 - System Dispatch: The AMGA generates an initial optimized plan using an Insertion Detection method to merge orders from the same restaurant to increase efficiency.
• Stage 2 - Rider Interaction: Riders receive instructions but can choose to accept only specific parts or suggest changes based on their immediate environment. This feedback is fed back into the system for a second-round local search optimization.

3. Order Merging Strategy

During peak hours, efficiency is won or lost in how orders are grouped. The paper introduces an Urgency Attribute. Orders with higher urgency (or those manually expedited by customers) form the "seed" for a task, and other orders from the same restaurant are merged into it using an insertion detection procedure.

Experiments and Insights

The study simulates a 5-km rectangular urban area with 30 customers and 5 restaurants. The output is visualized through a comprehensive Gantt Chart, which tracks rider loading, task allocation, and time windows.

Key Findings:

• Dynamic Adaptation: The AMGA's adaptive crossover and mutation rates allow it to jump out of local optima faster than standard GAs.
• Rider Empowerment: By allowing riders to have "certain autonomy," the system saw increased stability. When riders feel like participants rather than tools, the "happiness index" of the workforce increases, contributing to a more sustainable delivery ecosystem.

Critical Analysis & Conclusion

This paper makes a compelling case for moving away from "black-box" optimization toward Human-in-the-loop systems.

Strengths:

• Combines hard logistics optimization with soft human-centric factors.
• The two-layer coding simplifies the complex relationship between orders, restaurants, and riders.

Limitations:

• The model assumes a static set of orders at $T=0$. In reality, takeout is a "rolling horizon" problem where new orders arrive continuously.
• The HCI component requires high-quality mobile interfaces and reliable real-time data from riders, which may increase platform overhead.

Future Outlook: The next step for this research is to apply these HCI strategies to dynamic continuous-time periods, where the algorithm must re-optimize every few seconds as new orders "pop" onto the map.

Keywords: Takeout Delivery Problem, Bi-objective Optimization, HCI solution, Heuristic Algorithm, AMGA.