i-Rebalance: Personalized Vehicle Repositioning for Supply Demand Balance

Authors: Haoyang Chen, Peiyan Sun, Qiyuan Song, Wanyuan Wang, Weiwei Wu, Wencan Zhang, Guanyu Gao, Yan Lyu

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate the performance of i-Rebalance in a simulator built with a real-world taxi trajectory dataset. In the following, we first describe experimental settings and then discuss experimental results.
Researcher Affiliation Academia 1Southeast Univeristy, China 2National University of Singapore 3Nanjing University of Science and Technology {hy chen, spy, qiyuan song, wywang, weiweiwu}@seu.edu.cn, wencanz@u.nus.edu, gygao@njust.edu.cn, lvyanly@seu.edu.cn
Pseudocode Yes Algorithm 1: Personalized Taxi Repositioning
Open Source Code Yes The code for our approach is available at https://github.com/Haoyang-Chen/i Rebalance.
Open Datasets Yes We use a real-world taxi trajectory dataset collected in Chengdu, China, with a total of 14,865 taxis and 10,710,949 passenger trips within a month (Lyu et al. 2019).
Dataset Splits No The paper does not explicitly provide specific train/validation/test dataset splits (e.g., percentages or sample counts). It mentions using a real-world dataset but doesn't detail how it was portioned for training, validation, and testing.
Hardware Specification Yes The models are trained on Intel Core i9-10940K CPU @3.30GHz, NVIDIA Ge Force RTX 3090, and 32GB memory.
Software Dependencies No The paper mentions an "A2C network," "fully connected layers," "ReLU activations," "entropy loss," "batch normalization," and "Adam optimizer." However, it does not provide specific version numbers for any software frameworks (e.g., PyTorch, TensorFlow) or programming languages (e.g., Python).
Experiment Setup Yes In i-Rebalance, we implemented the actor and critic networks with two fully connected layers (64 units, Re LU activations), respectively. Grid Agent and Vehicle Agents use the same architecture, normalizing outputs with softmax and sigmoid functions separately. We used entropy loss and batch normalization, entropy-β is set to 0.01 and the batch size is 10. We set a decay factor γ = 0.98 and train the network with Adam optimizer (Liu et al. 2020). In reward function (Eq. 5), the weights αB = 2, and αP = 1 achieve the best performance.