Minimising Regret in Route Choice

The use of reinforcement learning (RL) in multiagent scenarios is challenging. I consider the route choice problem, where drivers must choose routes that minimise their travel times. Here, selfish RL-agents must adapt to each others' decisions. In this work, I show how the agents can learn (with performance guarantees) by minimising the regret associated with their decisions, thus achieving the User Equilibrium (UE). Considering the UE is inefficient from a global perspective, I also focus on bridging the gap between the UE and the system optimum. In contrast to previous approaches, this work drops any full knowledge assumption.