We model a ride-sharing platform (Uber/Lyft) as a Mean-Field Game (MFG). Each driver is a strategic agent who, at every minute, decides whether to accept or reject a randomly offered ride. Drivers don't reposition proactively — their zone changes only when they complete a trip.
We find the Mean-Field Nash Equilibrium: a joint policy from which no driver can unilaterally deviate to earn more, given all other drivers follow the same strategy.
Compute V(z,t) — expected future earnings — backwards from T→0. At each step, the driver compares Q_accept vs Q_reject using a sigmoid policy.
Run all drivers forward using the computed policy π. Drivers scatter to destination zones after accepting rides, producing a new µ_idle distribution.
When µ_new ≈ µ_old, we have found a fixed point. Every driver is already best-responding — no one can do better by deviating.
PoA = W_opt / W_MFNE measures welfare loss from selfish behavior. A PoA of 1.52 means drivers deliver only 66% of the socially optimal matching.
A 7×7 grid of 49 zones with synthetic demand. Three demand profiles test how spatial structure affects equilibrium behavior and the Price of Anarchy.
We start with every driver spread uniformly — 1/500 of drivers at each of the 49 zones at t=0. We can't just hold this uniform distribution for all T steps because by t=250, drivers have been moving for 250 steps and the real distribution looks nothing like uniform. Using uniform throughout would make the first backward pass compute p_offer values ~3–4× off at mid-horizon, and the outer loop never recovers. So we first run one forward simulation under an always-accept policy to get a physically consistent starting trajectory, then begin the fictitious play loop from there.
All three profiles converge with α = 0.05 and ε = 0.01. Log-linear L1 decay consistent with geometric convergence.
Single Gaussian demand peak at the center (3,3). Center zone is demand-saturated; edge zones rarely get offers. PoA ≈ 1.37.
Constant demand everywhere. All zones are equivalent so no spatial selectivity gradient exists. PoA ≈ 1.01 — selfish behavior is nearly optimal.
Two Gaussian demand peaks at (3,3) and (1,5). Drivers must choose which center to serve, slightly increasing strategic selectivity. PoA ≈ 1.40.
All synthetic assumptions replaced with parameters extracted from real trip records published by the NYC Taxi & Limousine Commission (TLC) — specifically the High-Volume For-Hire Vehicle (HVFHV) dataset covering Uber & Lyft rides. A plain Tuesday, 259 active TLC zones, full minute-level resolution.
Drivers queue at JFK (zone 132) for high-fare Manhattan runs (~$45–60), accepting longer waits. The TLC even maintains a dedicated driver holding lot here.
Phase 2 introduces age cohorts: a rejected ride re-enters the pool for up to 5 minutes before expiring. More realistic than the memoryless Phase 1 model.
Destination probabilities P_dest[z,z'] come directly from the data. Most JFK trips go to Midtown — not uniform across 259 zones.
Two differences from Phase 1. First, the starting distribution at t=0 is not uniform — drivers are placed proportional to total daily pickups per zone, so high-traffic zones like Midtown start with more drivers than quiet outer-borough zones. Second, Phase 2 also needs to initialize the 5-minute ride pool R[z,t] (the age-cohort backlog of unaccepted rides), not just µ. The always-accept warm start produces both together, giving the outer loop a physically consistent starting point for both the driver distribution and the ride pool.
Converged at iteration 166 with final L1 = 0.98. α=0.02 (smaller than Phase 1 to handle spiky real demand). ε=1.0 calibrated to the integer noise floor of real trip counts.
Real NYC demand is far spikier than synthetic demand, creating stronger incentives to cluster and widening the gap between selfish and optimal outcomes.
| Dimension | Phase 1 | Phase 2 |
|---|---|---|
| Zones | 49 synthetic (7×7 grid) | 259 real TLC zones |
| Time horizon | T = 500 steps | T = 1440 minutes (full day) |
| Demand X[z,t] | Gaussian hotspot (stationary) | Real trip counts (time-varying) |
| Fare model | γ · Manhattan distance | Median driver_pay from data |
| Trip duration | Manhattan distance / fixed speed | Median trip_time from data |
| Destination dist. | Uniform 1/(n−1) | Empirical O-D matrix P_dest |
| Ride persistence | Discard on reject | 5-minute pool (5 age cohorts) |
| Driver count N | 500 (chosen) | 10,979 (estimated from data) |
| α (smoothing) | 0.05 | 0.02 |
| ε (convergence) | 0.01 | 1.0 |
| Iterations | 59–139 | 166 |
| Busy fraction | 31.9% (hotspot) | 61.6% |
| Price of Anarchy | ≈ 1.37 (hotspot) | 1.52 |
Real NYC demand is spiky — airports and Midtown generate a disproportionate share of trips. This creates stronger clustering incentives, widening the gap between selfish and optimal matching.
Integer trip counts create a noise floor in the forward simulation. A larger α amplifies this noise causing oscillation. ε=1.0 is calibrated to this floor — it represents genuine convergence, not a loose threshold.
PoA = 1.52 means the selfish equilibrium delivers only 66% of the socially optimal ride matching. A central planner could match 52% more rides with the same number of drivers.
A way to find Nash equilibria when there are too many players to track individually.