Variational Poisson Recurrent Neural Network

Official implementation of Predictive and Prescriptive Performance of Bike-Sharing Demand Forecasts for Inventory Management

Prerequisites

To install all required dependencies, run

pip install -r requirements.txt

Contents

• src/algos/vprnn.py: PyTorch implementation of VP-RNN and MOVP-RNN modules (Section 3.2).
• src/algos/inventory_decision_hourly.py: implementation of queuing model for inventory management with 60-min interval (Section 3.1).
• src/algos/inventory_decision_quarterly.py: implementation of queuing model for inventory management with finer aggregation levels: 15-, 30-min intervals.
• src/misc/: helper functions.
• data/: csv files for NYC experiments.
• saved_files/: directory for saving results, logging, etc.
• run_vprnn.py: script for running VP-RNN on NYC data.
• run_movprnn.py: script for running MOVP-RNN on NYC data.

Examples

To run the VP-RNN pipeline, run_vprnn.py accepts the following arguments:

model arguments:
    --epochs        number of epochs to train (default: 50k)
    --no-cuda       disables CUDA training (default: False, i.e. run on CPU)
    --seed          random seed (default: 1)
    --patience      how many epochs without improvement to stop training
    --no-train      disables training process (default: False, i.e. train)
    --no-predict    disables prediction process (default: False, i.e. generate predictions)
    --no-decision   disables decision model (default: False, i.e. run queueing model)
    --benchmark     enables benchmark decision (default: False, i.e. disables benchmark decision calculation)
    
    
data arguments:
    --stations       list of station IDs on which to run pipeline
    --interval       defines temporal aggregation (defaul: 60min)

Training, generating demand predictions & inventory decisions

1. To train, compute demand predictions, and inventory decisions with an VPRNN model (using the default parameters) on e.g. station 426 and 229, run the following:

python run_vprnn.py --stations 426 229

During execution, this will store the following:

• Model parameter after training in saved_files/{15,30,60}min/trained_models/rnn/
• Prediction and Optimization performance in saved_files/{15,30,60}min/results/
• Predicted demand vectors in saved_files/{15,30,60}min/predicted_demand/
• Qualitative visualization of demand predictions in saved_files/{15,30,60}min/images/
• Inventory decision in saved_files/{15,30,60}min/inventory_decisions/

2. To only generate inventory decisions with different pre-trained demand predictions for comparison, on e.g. station 426 and 229, run the following:

python run_vprnn.py --stations 426 229 --no-train 1 --no-predict 1 --benchmark

Note that the inventory decision model needs to read pre-trained demand predictions stored in saved_files/{15,30,60}min/predicted_demand/. Make sure there are predicted demand vectors stored in this directory.

During execution, this will store the following:

• Inventory decision in saved_files/{15,30,60}min/inventory_decisions/
• Optimization performance in saved_files/{15,30,60}min/results/optimization

Credits

This work was conducted as a joint effort with Yihua Wang*, Dennis Prak*, Filipe Rodrigues', Stefan Minner* and Francisco C. Pereira', at Technical University of Denmark' and Technical University of Munich*.

Reference

@inproceedings{GammelliEtAl2021,
  author = {Gammelli, D. and Wang, Y. and Prak, D. and Rodrigues, F. and Minner, S. and Pereira, F. C.},
  title = {Predictive and Prescriptive Performance of Bike-Sharing Demand Forecasts for Inventory Management},
  year = {2021},
  note = {Submitted},
}

---

In case of any questions, bugs, suggestions or improvements, please feel free to contact me at daga@dtu.dk.