In real traffic networks, collective autonomous vehicle (AV) route choices create externalities for human‑driven vehicles (HDV); routes that benefit AVs can increase travel times for others. To mitigate this, we augment the selfish travel‑time reward with a social component called marginal cost. Marginal cost measures an AV’s impact on other users by comparing their travel times with those from a counterfactual simulation where the AV is removed.
We employ multi‑agent reinforcement learning (MARL) in mixed‑autonomy settings, allowing AVs to learn routing policies through repeated interaction with the traffic environment. Experiments on the real‑world networks of Ingolstadt and Provins show that adding marginal cost shortens AV travel times relative to selfish routing and reduces HDV travel‑time growth from 1.7% to 1.1%. Overall system performance improves and AVs also benefit. This pattern holds across cities, penetration rates, and repetitions, but excessive weight on the social term prevents the system from reaching desired performance.
For scalability, we compute marginal cost only for AVs whose routes most affect traffic, cutting the number of counterfactual simulations while retaining similar performance at much lower computational cost. The results suggest AV routing should move beyond selfish rewards and incorporate a social component to benefit all participants.
Review: Introducing marginal cost into the reward function provides a principled way to balance efficiency and fairness, and its effectiveness across diverse scenarios makes it a promising direction for large‑scale deployment.