Operating Bike-Sharing Systems – A Case Study in Real-Time Decision Making
summary
Citi Bike data-driven work models stations as time-varying Poisson flows and computes a user dissatisfaction function (expected stock-outs) from 30-minute Markov models after de‑censoring demand; real‑time tools produce greedy truck routes and prefix plans but total motorized rebalancing capacity is tiny so trikes for rebalancing (6–12 bikes, 6‑minute loops) and overnight box trucks (≈50 bikes) are used. A one‑time reconfiguration moving ~500–1,600 docks yields ~1,200 fewer outages (~30%); an incentive pilot called Bike Angels (lottery prizes) shifts many riders to helpful stations.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 5:05Demo
- 7:22Analysis
- 12:58Missing Data
- 24:31Performance
- 27:53Cost Function
- 31:07Scatter Plots
- 35:34Midrash Rebalancing
- 37:38Offline Decision Making
- 41:47Overnight Rebalancing
- 43:25RealTime Decision Making
- 46:33Bouncing
- 48:06System Design
- 53:53Stochastic Model
- 54:50Experimental Design
- 56:11Experimental Results
related talks
34mData Drives Business
06mDynamic Public Resource Allocation Based on Human Mobility Prediction
19mThe Data Science behind Cruise’s Robotaxi Fleet
31mScaling up Lyft’s Growth via Causal ML and Bandits
29mIncentives and Games in Large Scale Mobility
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
48mUrban Data & Decisions
34mMove Your Crate - How to optimize your mobility