Verification of thevenin's theorem for BEEE Lab (1).pptx
Architecting a real time optimization platform for driver positioning (informs 2019)
1. Architecting a real-time
optimization framework for
Driver Positioning
Hao Yi Ong
w/ many others at Lyft
INFORMS Annual Meeting | Seattle | October 22, 2019
5. Personal Power Zones (PPZ) product
Pre-empt general market imbalance
and help with events or demand
spikes
Algorithmically coordinate drivers
when whole cities light up with rider
demand
De-conflicting drivers Early signaling
10. PT problem
The PT model maps market conditions (supply and demand) to the general notion of value.
// value; e.g., rides
// match supply/demand
// limit price
fixed parameter:
treatable supply
untreatable supply
price-adjusted
demand
reserved supply
11. PT problem
The PT model maps market conditions (supply and demand) to the general notion of value.
MKT model
demand by gh6
supply by gh6
value
12. PPZ problem
Supply and supply allocation are the PPZ decision variables (in addition to PT). The PPZ
model repositions supply to maximize value.
PPZ model
MKT model
forecasted demand by gh6
supply by gh6
value
iteratively refine
supply allocation
“feedback”
13. PPZ-induced value
// move supply
// allocation constraints
decision variable:
treatable supply
PPZ problem
Supply and supply allocation are the PPZ decision variables (in addition to PT). The PPZ
model repositions supply to maximize value. (Omitted: various other business constraints.)
decision variable:
supply allocation
initial treatable
supply
15. Assumptions
● Drivers don’t reposition themselves for decision look-ahead
● Driver compliance exponentially decays as travel time increases (fitted)
● When drivers comply, they move to destination by the next time step
Methodology
● Historically accurate data importance-sampled over a quarter over top-10 cities
● “Fractional drivers” rounded down to the nearest integer to compute “actual value”
Results (varies by city)
● 1—4% incremental rides
● 0.2—1% incremental bookings
Numerical experiments
16. Methodology
● Time-split experiments infeasible due to product experience
● User-split testing is the least bad option
● Impact estimates derived from causal inference methods
Results (varies by city)
● ~0.3% incremental bookings
● Overall, drivers prefer PPZ over PT
● PPZ rolled out to virtually everywhere
User-split experiments