PROJECT TITLE :

Deployment Optimization for Shared e-Mobility Systems with Multi-agent Deep Neural Search

ABSTRACT:

Shared e-mobility services have undergone extensive testing and pilot programs in cities all over the world, and they are already integrated into the framework of contemporary urban planning. This paper investigates a problem in those systems that is both practical and important: the question of how to deploy and manage their infrastructure across space and time in such a way that the services can be ubiquitous to the users while also remaining profitable. However, in real-world systems, evaluating the performance of various deployment strategies and then finding the optimal plan is prohibitively expensive. This is because it is frequently not feasible to conduct many iterations of trial-and-error. In order to find the optimal plan, it is necessary to evaluate the performance of the different deployment strategies. We solve this problem by developing a high-fidelity simulation environment, which abstracts the key operation details of the shared e-mobility systems at a fine-grained level, and which is calibrated using data collected from the real-world environment. Before actually putting any of them into practice in the real-world systems, this enables us to experiment with any deployment plan we choose in order to determine which is the most effective given the particular circumstances. In particular, we propose a brand new multi-agent neural search approach, in which we design a hierarchical controller to produce tentative deployment plans. This approach uses neural networks to search for patterns across multiple agents simultaneously. After the deployment plans have been generated, they are tested with a multi-simulation paradigm, which means they are evaluated in parallel. The results of these tests are then used to train the controller with deep reinforcement learning. Because of this closed loop, the controller can be steered to have a higher probability of generating better deployment plans in subsequent iterations. This will allow for greater flexibility. The proposed method has been rigorously tested in our simulation environment, and the results of the experiments show that it is superior to baselines such as human knowledge as well as state-of-the-art heuristic-based optimization approaches with regard to service coverage as well as net revenue.