Rebalancing Expanding EV Sharing Systems with Deep Reinforcement Learning

Authors: Man Luo, Wenzhe Zhang, Tianyou Song, Kun Li, Hongming Zhu, Bowen Du, Hongkai Wen

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed approach using a simulator calibrated with 1-year operation data from a real EV sharing system. Results show that our approach significantly outperforms the state-of-the-art, offering up to 14% gain in order satisfied rate and 12% increase in net revenue. We build a simulator which is calibrated with 12 months operation data collected from a real-world EV sharing system . The proposed approach has been evaluated extensively, and results show that it significantly outperforms the state-of-the-art, offering up to 12% improvement in net revenue and 14% in demand satisfied rate.
Researcher Affiliation Collaboration Man Luo1,2 , Wenzhe Zhang3 , Tianyou Song3 , Kun Li3 , Hongming Zhu3 , Bowen Du1 and Hongkai Wen1 1Department of Computer Science, University of Warwick, UK 2The Alan Turing Institute, UK 3School of Software Engineering, Tongji University, China
Pseudocode No The paper describes the proposed approach (e.g., in Section 3.2 "Policy Optimization with Action Cascading") but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code available at https://github.com/ev-sharing/simulator.
Open Datasets No We build a simulator which is calibrated with 12 months operation data collected from a real-world EV sharing system . The paper states the data is from a "real-world EV sharing system" in Shanghai, but provides no link or citation for public access to this specific dataset.
Dataset Splits No The paper discusses using a simulator calibrated with real-world data and training a neural network. However, it does not provide specific details on how the dataset (simulated or real) is split into training, validation, and test sets for the MARL algorithm itself.
Hardware Specification Yes All the competing approaches are implemented with Tensor Flow 1.14.0, and trained with a single NVIDIA 2080Ti GPU.
Software Dependencies Yes All the competing approaches are implemented with Tensor Flow 1.14.0, and trained with a single NVIDIA 2080Ti GPU.
Experiment Setup Yes The weights α1, α2 and α3 scale different reward/penalty terms to approximately the same range, which are determined empirically via grid search. ϵ is a hyperparameter (usually set to 0.2 0.3).