Multi-Agent Market Effect of AI Policy Adoption

Research question: Does an AI repositioning recommendation degrade when an increasing share of drivers follows the same policy?

Answer (simulation): Yes. Above ~50% adoption the AI-recommended cohort’s revenue advantage disappears and turns negative, while market saturation rises toward 90%+ of pickup attempts.

Evaluation type: SIMULATION (finite-demand multi-agent simulator). Not a real-world market, deployment, or A/B result.

Experiment

  • Simulator: src/simulator/multi_agent/engine.py β€” finite-demand, explicit competition, every trip assignable at most once.

  • Fleet: 100 drivers, 7 simulated days, seed 42.

  • Policy: Two-Step Horizon (src/2_recommendation_algorithm/improved_strategy.py) for the AI cohort; baseline drivers follow the Hot Zone heuristic.

  • Adoption rates: 1%, 5%, 10%, 25%, 50%, 75%, 100%.

  • Artifact: outputs/experiments/adoption_sweep.json (regenerate via python scripts/run_adoption_sweep.py).

Results

Adoption

AI revenue/driver

Baseline revenue/driver

Revenue gap

AI utilization

Zone concentration (top-3)

Saturation

1%

$24.98

$21.09

+$3.89

18.2%

57.1%

0.0%

5%

$27.06

$21.21

+$5.85

17.9%

31.4%

0.0%

10%

$24.64

$21.17

+$3.47

15.5%

22.9%

0.9%

25%

$23.17

$21.29

+$1.88

15.3%

24.0%

43.0%

50%

$21.24

$21.90

βˆ’$0.65

14.8%

19.4%

74.5%

75%

$19.20

$20.93

βˆ’$1.73

14.2%

19.8%

83.6%

100%

$17.24

β€”

β€”

13.1%

18.6%

92.1%

At 100% adoption there is no non-AI cohort left to compare against, so the baseline column is empty.

Interpretation

  1. The advantage is fragile. At low adoption (1–25%) the AI cohort earns more than the heuristic cohort. Above 50% the gap flips negative: the baseline cohort now earns more than the AI cohort.

  2. The mechanism is saturation. saturation_rate β€” the fraction of pickup attempts in zone-slots where competing supply exceeded remaining trip inventory β€” rises monotonically from 0% to 92.1%. When many drivers follow the same recommendation, they converge on the same zones, compete for a finite pool of trips, and cannibalize each other.

  3. Concentration is not the whole story. Zone top-3 concentration actually falls as adoption rises (57% β†’ 19%). The AI fleet spreads out, but total demand is fixed, so spreading does not create new trips β€” it just dilutes the per-driver share of a fixed pool.

  4. Implication for deployment. A single shared policy has a self-defeating equilibrium. Production systems need either stochastic policies (exploration), fleet-level coordination that internalizes competition, or heterogeneous policies β€” none of which are tested here.

Limitations (read before citing)

  • Single seed (42), single fleet size (100), 7-day horizon.

  • No driver adaptation, learning, or abandonment.

  • No congestion coupling (this experiment holds travel times fixed).

  • Revenue figures are simulator proxies, not earnings.

Follow-up directions

  • Sweep fleet size and seed to confirm the crossover point (~50%).

  • Vary policy diversity: how much heterogeneity restores the advantage?

  • Add congestion coupling (src/simulator/v2/dynamics.py traffic multiplier) and test whether concentration now increases travel time.

  • Compare a stochastic (epsilon-greedy) version of Two-Step vs the deterministic one.