Waymo · Statistics & Data Analysis
Compare Driving Algorithms Using Hard-Braking Rates
TrueInterview
October 7, 2026 · 2 min read
You need to compare two autonomous-driving algorithms. The table you have contains algorithm, city, road-condition category, miles driven, and number of hard-braking events. In this exercise, comfort is defined solely by hard braking: with similar driving exposure, fewer hard-braking events mean a more comfortable ride. Design an experiment and a statistical analysis for comparing the algorithms. Explain how mileage, city, road conditions, and the distribution of hard-braking counts influence your design, and point out what the existing table cannot establish on its own.
Constraints and Clarifying Questions
- Clarify what a single row represents: one trip, one vehicle over a period, or an aggregate of several drives.
- Define hard braking consistently for both algorithms, and decide whether miles are a comparable exposure measure across different road conditions.
- Clarify whether the algorithms were assigned at random and whether the same vehicles or routes produce repeated observations.
- No sample size, count distribution, improvement target, or assignment mechanism is given, so state any statistical model as a hypothesis to be checked.
- The task is to design and analyze an experiment; a database query by itself is not an answer.
Hint: Keep exposure together with the count. Two drives can record the same number of hard-braking events yet cover very different mileage. Also think about what happens when one algorithm gets most of the difficult-road mileage.
What a Strong Answer Covers
- A comfort metric, an estimand, and a comparison direction that account for driving exposure.
- An assignment plan and analysis that handle city and road-condition differences without conflating stratification with randomization.
- A suitable starting model for count data, checks for overdispersion or excess zeros, and inference that accounts for repeated observations.
- An effect estimate and uncertainty interval, plus the limitations of observational or aggregated data.
Follow-up Questions
- Under what conditions would a simple Poisson model be inappropriate for these counts?
- How would your conclusion change if one algorithm were tested mostly under easier road conditions?
- What extra identifiers would you collect if rows from the same vehicle were correlated?
- How would you design the experiment when hard-braking events are rare and many rows have zero events?
Overview: Design an exposure-aware experiment that compares hard-braking rates between two driving algorithms, with controls for city and road conditions and count-data inference.