Route-Shopper Pairing With Parallel Candidate Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Online concierge systems face high latency and scalability issues in selecting delivery agents to fulfill orders due to the large number of combinations of retail locations and shoppers involved, exacerbated during peak hours and with increasing popularity.
Innovation Solution
An asynchronous process is used to select optimal pairings of delivery agents with routes, involving parallel candidate generation and optimization processes to reduce latency and improve scalability, by partitioning routes into sets for parallel processing and using a cost engine to determine efficient pairings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the system uses traditional sequential processes to select shoppers for orders, then it can consider all constraints and factors comprehensively, but the latency increases significantly and scalability is poor
Solution Approach 1:
The patent segments the shopper selection process into multiple independent worker processes that operate in parallel. Each worker handles a subset of orders and generates candidate pairings independently, allowing the system to process many orders simultaneously rather than sequentially, thus reducing overall latency while maintaining comprehensive constraint checking within each worker's subset
Solution Approach 2:
The system performs preliminary actions by pre-generating candidate pairings for each order with multiple shoppers before optimization begins. This preliminary candidate generation phase prepares multiple potential solutions in advance, allowing the optimization process to quickly select the best pairing without having to evaluate all possible combinations from scratch, thereby reducing total processing time
2Productivity
If the system increases processing power to handle more orders simultaneously, then latency decreases, but the complexity of managing parallel processes and resources increases
Solution Approach 1:
The patent implements dynamic process management where worker processes are spawned and terminated based on the current workload. The system dynamically adjusts the number of parallel workers and their resource allocation according to the number of orders needing fulfillment, allowing high productivity during peak times while reducing complexity during low-demand periods
Solution Approach 2:
The system introduces an intermediary optimization process that coordinates between the parallel worker processes and the final selection. This intermediary layer manages the candidate pairings generated by multiple workers, applies optimization algorithms, and produces the final shopper assignments, thereby simplifying the overall system architecture while enabling parallel processing of many orders
3Reliability
If the system generates and evaluates all possible shopper-order pairings, then optimal pairings can be selected, but the computational time and resources required become excessive
Solution Approach 1:
The system applies partial action by generating candidate pairings for only a subset of shoppers for each order rather than evaluating all possible shoppers. The candidate generation phase creates a limited set of promising pairings based on initial filtering criteria, and the optimization process then selects the best among these candidates, achieving high-quality matching without the excessive computational cost of evaluating all possible pairings
Solution Approach 2:
The patent changes parameters by transforming the shopper selection problem from evaluating all possible pairings to working with candidate pairings that have pre-filtered parameters. The system modifies the search space by applying constraints and filtering criteria to generate only relevant candidate pairings, thereby reducing the computational complexity while maintaining the ability to find optimal or near-optimal solutions
Data Source
AI summary
An online concierge system receives information describing orders from its customers and generates a route for each order based on this information. The routes are partitioned into multiple sets of routes and multiple candidate generation processes are executed in parallel. During execution of a candidate generation process, one or more routes included in each set of routes are paired with shoppers of the system based on a set of constraints, producing multiple route-shopper pairs. A cost associated with each route-shopper pair is determined based on attributes associated with each shopper and/or information associated with each route of the pair. During an optimization process, which is executed asynchronously with the candidate generation process, one or more route-shopper pairs are selected based on pairing-cost data identifying route-shopper pairs and their associated costs. One or more requests to fulfill orders are sent to one or more shoppers based on the selected route-shopper pair(s).