Optimal Online Dispatch for High-Capacity Shared Autonomous Mobility-on-Demand Systems

A Value-based Dynamic Learning Approach for Vehicle Dispatch in Ride-Sharing

“…Online Dispatch. The dispatcher uses the candidate generation method in [2] to compute actions for all vehicles. It then employs the learned value function to match vehicles and requests in a far-sighted way, by scoring candidate matches (i.e., candidate schedules) with the sum of instant rewards and long-term expectations.…”

“…To enable high occupancy ride-sharing, a scalable batch assignment algorithm is developed in [1] to utilize the full capacities of vehicles. The efficiency of the fleet is further improved in [2] by exploring all possible ride-sharing combinations in real time and dynamically altering the assignment policy to keep it "optimal" at any given time. However, to provide a good passenger experience, the above approaches involve using small batch periods and are myopic in nature.…”

“…However, in practice, the decisions have to be made at t before the realization of the exogenous information w t+1 (arriving between t and t + 1), so as not to affect the passenger experience. Therefore, a post-decision state s a t [14] (also called afterstate [15]) is introduced to split the transition function into two components: s a t = f (1) (s t , a t ) and s t+1 = f (2) (s a t , w t+1 ). The post-decision state is a segregation of deterministic and stochastic information about the future and explicitly captures the most recent state.…”

“…The most common approach is batch assignment, which uses dispatch windows to allocate multiple requests at the same time. It is well suited for real-world ride-sharing production systems and enables a "global" optimization for each dispatch window [1], [2]. However, because of the lack of future information in the subsequent batch windows, an assignment decision that is "optimal" now may cause potential efficiency losses in the future and leaves room for further optimizations [3], [4].…”

“…2) We adopt an online learning procedure to continuously update the offline learned value function to handle nonstationary dynamics, using only the states of vehicles in the current operating epoch. 3) We combine the learned value function with the reoptimization dispatch method in [2] to optimize the objective over a long horizon, including both recent and upcoming epochs. 4) We conduct simulation experiments on real world largescale taxi datasets to evaluate our proposed approach and analyze the impact of long-term optimization.…”