Dynamic incentives for efficient traffic management: a reinforcement learning approach

Traffic management literature often overlooks encouragement-based strategies in favour of road pricing. However, road pricing often raises concerns about accessibility and public dissatisfaction, leading to its prohibition in some places such as Finland. Similar to pricing, incentives can push the User-Equilibrium (UE) flow pattern towards the System Optimum (SO). We formulate and solve a problem to determine the allocation of incentives, encouraging users to reroute onto an alternative (potentially longer) path, to achieve overall social benefit. The proposed approach uses multi-agent reinforcement learning to assign incentives to drivers, in which travel times are estimated dynamically using a traffic simulation software (SUMO). We also introduce a dynamic pseudo-SO as a benchmark for evaluating the incentives’ effectiveness. Under an unlimited budget, the incentives scheme achieves virtually identical performance to the dynamic pseudo-SO. As the budget decreases, the solution gradually degrades, reaching the dynamic UE. The numerical results demonstrate that unlimited incentives can reduce the total travel time by an average of 23%. However, employing a budget time equivalent to around 6.3% of the UE total travel time can achieve a 10% reduction in total travel time.