Selective Machine-Learned Model Distribution for Ride-Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computational intensive processes for generating predictions in ride-sharing systems using machine-learned models are inefficient due to the large size of models and extensive input data, making it difficult to provide timely predictions like estimated time of arrival (ETA) for riders.
Innovation Solution
A method that identifies a hierarchy of parent-child relationships based on travel information, trains machine-learned models, and selectively distributes them to computer servers, ensuring each server farm receives only the most accurate models relevant to its geographic area, reducing computational load and improving efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If machine-learned models are used to generate predictions quickly, then prediction speed is improved, but computational resources and model size increase
Solution Approach 1:
The patent segments the monolithic machine-learned model into multiple specialized sub-models, each trained for specific geographic regions or trip conditions. This segmentation allows the system to load and execute only the relevant sub-model for each prediction task, reducing memory footprint and computational overhead while maintaining prediction speed.
Solution Approach 2:
The patent implements local quality by creating region-specific or condition-specific sub-models tailored to local characteristics (e.g., urban vs. suburban areas, different times of day). Each sub-model is optimized for its specific context, improving prediction accuracy for local conditions while allowing the system to select only the necessary sub-models for each prediction, thereby reducing overall computational resource requirements.
2Measurement precision
If large machine-learned models are used to improve prediction accuracy, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent divides a large complex model into multiple smaller specialized sub-models, each optimized for specific geographic regions or trip conditions. This segmentation maintains high prediction accuracy for each local context while significantly reducing the memory and computational resources required compared to deploying a single large model for all scenarios.
Solution Approach 2:
The patent changes the parameters of the machine-learned models by training separate sub-models with different parameter sets for different geographic regions and conditions. Each sub-model has parameters optimized for its specific context, achieving high local accuracy while the system overall uses fewer resources by selecting only the relevant sub-models needed for each prediction task.
Data Source
AI summary
Machine-learned models are selectively distributed to a plurality of computer servers according to conditions associated with the computer servers. A server receives travel information from a travel coordination system. The travel information describes a plurality of conditions. The server identifies a hierarchy of one or more parent-child relationships based on the plurality of conditions. The server trains machine-learned models using the plurality of conditions described by the travel information. The server selects machine-learned models for the plurality of conditions responsive to the identified hierarchy. The server distributes machine-learned models to the plurality of computer servers responsive to the identified hierarchy.