Mixture of On-Demand and Pre-Booked Ride Sharing System

Article:

Journal Version: Zijian Zhao, Jing Gao*, Sen Li, "Real-Time Order Assignment for Ride-Sharing Platforms with a Mixture of Pre-booked and On-Demand Requests" (under review)

Conference Version: Zijian Zhao, Jing Gao*, Sen Li, "Ride-Hailing Order Dispatching with A Mixture of On-Demand and Pre-Booked Requests via Reinforcement Learning", 2026 COTA International Conference of Transportation Professionals (CICTP 2026)

1. Workflow

2. Preparation

2.1 Dataset

In our paper, we utilize the yellow taxi data from July 1, 2024, between 7:00 and 8:00 AM in Manhattan, sourced from TLC Trip Record Data - TLC. Both the raw data and our processed results are stored in the ./data folder. You can easily modify the time settings in main.py to train or evaluate the model for different time periods. However, if you wish to use data from other districts, you'll need to preprocess it according to the steps outlined in ./data/Manhattan_dic.py. Additionally, you will need to adjust the norm function in Worker.py to set the appropriate range for latitude and longitude. While it's possible to skip normalization, doing so may negatively impact performance.

The processed data can be found in the ./data directory.

Considering the copyright, we have removed the processed data. However, the data processing code is available in the ./data directory. Please download the dataset from the link provided above and use our code to process it.

2.2 OSRM

Before executing the code, you must first start the Docker container for OSRM (specifically for the northeastern region of the US). To prevent any port conflicts with our other projects, we will use port 6000 instead of the default:

docker run -t -i -p 6000:6000 -v "${PWD}:/data" ghcr.io/project-osrm/osrm-backend osrm-routed --algorithm mld /data/us-northeast-latest.osrm -p 6000

3. How to Run

3.1 Train

3.1.1 Train for a General Strategy

In this mode, the proportions of pre-booked orders, pooling orders, and the appearance times of pre-booked orders are randomized in each episode. This approach helps develop a robust strategy with minimal performance sacrifice.

python train.py --bi_direction --rand_rate --rand_appear

3.1.2 Train for a Specific Strategy

To achieve optimal performance for specific conditions, you can train the model using the following command:

python train.py --advance_time <time when pre-booked orders start to appear in adavance> --pooling_rate <proportion of customers choosing pooling> --prebooked_rate <proportion of pre-booked orders in total orders>

It would be beneficial to first pre-train a general strategy and then fine-tune the model for specific scenarios. To do this, you should load the pre-trained model parameters using --model_path <pre-trained model path>.

3.2 Evaluate

python eval.py --bi_direction --model_path <trained model path> --prebooked_rate <a list of your evaluation scenarios> --pooling_rate <a list of your evaluation scenarios> --advance_time <a list of your evaluation scenarios>

4. Citation