Multi-agent reinforcement learning for optimization of mixed autonomy traffic at scale
While the promised appearance of fully autonomous vehicles has been pushed back further and further, our highways have silently been transformed by the increasing penetration of hands-off adaptive cruise controllers. We investigate how, given current levels of cruise control availability, we can design driving strategies for these cruise controllers that increase the energy efficiency of the highway by smoothing out spontaneously forming shockwaves. Using multi-agent reinforcement learning, we show that we can design controllers that approximately act like they know the equilibrium speed of the system. These controllers outperform hand-designed control strategies and are robust to variations of the underlying dynamics.
Password: “arizona” (all lower case)