Using Machine Learning and Optimization to reallocate the docks across 70 bike sharing station in SF Bay Area. This project was completed as part of an MIT class.
Data
The dataset for this project is publicly available here: https://www.kaggle.com/benhamner/sf-bay-area-bike-share
Method
The analysis is split in two steps:
- Predict the demand for the next month using Optimal Regression Trees (Bertsimas et al., 2017). This results in the following tree (truncated for demonstrative purposes).
- Optimize the dock allocation to minimize the number of days with low stock of free docks using a Mixed Integer Optimization (MIO) formulation. The program is optimized to solve in under 3 seconds. This results in the following suggested changes to the dock allocation in downtown SF. On the left is the old allocation, on the right in red - changes number of docks at stations, in blue - unchanged. The size of the circles corresponds to the number of docks.
Languages
The data is cleaning in done in R. The predictive model as well as the optimization is done in Julia with Gurobi as the solver.