Explore a machine learning-based metaheuristic approach for efficiently reoptimizing autonomous ridesharing plans in dynamic environments. Delve into the local search-based metaheuristic that utilizes destroy-repair operators selected through machine learning, trained on over 1.5 million examples of solved ridesharing subproblems. Examine computational experiments conducted on dynamic instances from Uber Technologies Inc. data, showcasing the proposed approach's 9% average performance improvement over benchmark data-driven metaheuristics. Gain insights into the correlation between vehicle routing features and metaheuristic performance in autonomous ridesharing operations, presented by Claudia Bongiovanni from HEC Montréal in this 27-minute DS4DM Coffee Talk at GERAD Research Center.
Guiding Metaheuristics Through Machine Learning Predictions for Dynamic Autonomous Ridesharing
GERAD Research Center via YouTube
Overview
Syllabus
Guiding metaheuristics trough machine learning predictions
Urban Mobility and Logistics
The Dial-a-Ride Problem (DARP)¹
The Electric Autonomous Dial-a-Ride Problem (e-ADARP)2
The (Dynamic) e-ADARP
Two-Phase Metaheuristic
Popular Operators and Metaheuristics
Machine Learning-Based Large Neighborhood Search
The Uber Dataset?
Event-Based Simulation Framework
Creating Examples (Labeled Dataset)
Statistics Example
Extracted Features
The Prediction Problem
The MLNS Algorithm
ML: Training Phase
ML: Performance Measures
ML: Features Importance
Optimization: Validation Phase
Expected difference in the objective function improvemen
Summary of Contributions
Taught by
GERAD Research Center
