Graph based sparse neural networks for traffic signal optimization

Summary

This repository contains companion code for the paper Graph based sparse neural networks for traffic signal optimization, submitted for NeurIPS Conference 2019.

Trained models from our experiments can be found in the file 100k_fit_and_eval_experiment_results.zip that can be obtained from https://drive.google.com/file/d/1zch1oqMd_VEPSqm42I8njT2cJ9mN96Xi/view?usp=sharing and should be unzipped to the repository main directory.

The core dataset can be downloaded from https://drive.google.com/file/d/1UHGcHeJrxskkJ7ib9oe9oSz21nRnrCVN/view?usp=sharing and should be saved in the repository main directory.

The toy dataset can be found in the archive toy_set.zip, available from https://drive.google.com/file/d/1y-25uIiPRQb7zUNQshrh8UXtegLKg25e/view?usp=sharing. The archive should be unzipped to the repository main directory.

Repository contents description

Core modules:

• graph_neural_networks.py contains basic methods for constructing graph neural networks. This is the core of the whole repository
• graph_utils.py contains several utility methods for working with graphs
• data_preparation_utils.py contains simple methods for preparing data for training (including scaler initialization)
• training_and_evaluation.py contains methods for training and evaluation of models, reused many times in our experiments
• iterative_updaters.py contains a few basic iterative updaters for use in gradient descent experiments
• toy_examples.py currently contains core code for a single toy problem, related to neighbourhood entropy (similar to (2) in t-SNE paper (van der Maaten, L. and Hinton, G. (2008)))

Experiment scripts:

• 100k_fit_experiment.py contains code for running main experiment with multiple graph nn architectures of type 1 (Table 1 in the paper). Fits models.
• 100k_eval_experiment.py contains code for running main experiment with multiple graph nn architectures of type 1 (Table 1 in the paper). Evaluates models produced by 100k_fit_experiment.py.
• 100k_fit_experiment_kipf.py contains code for running main experiment with multiple graph nn architectures of type 2 (Table 2 in paper). Fits models.
• 100k_eval_experiment_kipf.py contains code for running main experiment with multiple graph nn architectures of type 2 (Table 1 in paper). Evaluates models produced by 100k_fit_experiment_kipf.py.
• 100k_fit_experiment_feedforward.py contains code for running reference experiment with a few feedforward architectures (Table 3 in paper). Fits models.
• 100k_eval_experiment_feedforward.py contains code for running reference experiment with a few feedforward architectures (Table 3 in paper). Evaluates models produced by 100k_fit_experiment_feedforward.py.
• random_topologies_experiment.py contains code for running random graph experiments for graph nns of type 1 (method 1, Figure 1(a) in the paper).
• permuted_topologies_experiment.py contains code for running random (permuted) graph experiments for graph nns of type 1 (method 2, Figure 1(b) in the paper).
• random_topologies_experiment_kipf.py is a analogue of random_topologies_experiment.py for graph nns of type 2 (check Supplementary materials for results)
• permuted_topologies_experiment_kipf.py is a analogue of permuted_topologies_experiment.py for graph nns of type 2 (check Supplementary materials for results)
• toy_random_topologies_experiment.py is an analogue of random_topologies_experiment.py for our toy problem (check Supplementary materials for results)
• toy_permuted_topologies_experiment.py is an analogue of permuted_topologies_experiment.py for our toy problem (check Supplementary materials for results)

Notebooks:

• graph_nn_type_1_tests.ipynb is a notebook documenting a number of experiments related to type 1 graph nns, including preliminary results, random graph experiment summary, and some nn visualization (not included in the paper)
• graph_nn_type_2_tests.ipynb is an analogue of graph_nn_type_1_tests.ipynb for type 2 networks.
• feedforward_nn_tests.ipynb contains first test results for feedforward neural networks, in preparation to produce Table 3 in the paper.
• fit_and_evaluate_experiments_summary_type_1.ipynb was used for producing Table 1 in the paper, as well as two tables included in Supplementary materials.
• fit_and_evaluate_experiments_summary_type_2.ipynb is a full analogue of fit_and_evaluate_experiments_summary_type_1.ipynb for graph nns of type 2. Table 2 was obtained using this code, as well two further tables included in Supplementary materials.
• fit_and_evaluate_experiments_summary_feedforward.ipynb is a full analogue of fit_and_evaluate_experiments_summary_type_1.ipynb for fully connected feedforward networks. Table 3 was obtained using this code, as well two further tables included in Supplementary materials.

Result files:

• 100k_fit_and_evaluate_experiments/fit_eval_results.csv contains basic results of the main experiment for type 1 graph nns. These results are now part of Table 1 in the paper.
• 100k_fit_and_evaluate_experiments_kipf/fit_eval_results_kipf.csv is an analogue of 100k_fit_and_evaluate_experiments/fit_eval_results.csv for type 2 networks. These results are now part of Table 2 in the paper.
• 100k_feedforward_fit_and_evaluate_experiments/fit_eval_results.csv is an analogue of 100k_fit_and_evaluate_experiments/fit_eval_results.csv for fully connected feedforward nns. These results are now part of Table 3 in the paper.
• random_topologies_3_2.csv contains results of random graph experiments for type 1 graph nns (method 1, Figure 1(a) in the paper)
• permuted_topologies_0.csv contains results of random graph experiments for type 1 graph nns (method 2, Figure 1(b) in the paper)
• random_topologies_3_2_kipf.csv is an analogue of random_topologies_3_2.csv for type 2 networks. Used for producing a plot in the Supplementary materials for the paper.
• permuted_topologies_0_kipf.csv is an analogue of permuted_topologies_0.csv for type 2 networks. Used for producing a plot in the Supplementary materials for the paper.
• toy_random_topologies_0.csv contains results of random graph experiments for type 1 graph nns based on our toy problem (graph randomization method 1). Used for producing a plot in the Supplementary materials for the paper.
• toy_permuted_topologies_0.csv contains results of random graph experiments for type 1 graph nns based on our toy problem (graph randomization method 2). Used for producing a plot in the Supplementary materials for the paper.

Additional:

• macierz_sasiedztwa.txt contains our traffic optimization problem adjacency matrix. It is a directed, nonsymmetric matrix, but we symmetrize it in all our experiments by extracting undirected edges. "Macierz sąsiedztwa" just means "adjacency matrix" in Polish.
• environment.txt contains the list of python packages in the Python environment we used on AWS.
• PNG files contain the plots included in the paper and its Supplementary materials.

NOTE: Type 2 neural networks were first dubbed "Kipf's networks" by us due to a formal analogy to the networks discussed in (Kipf, T. N. and Welling, M. (2017)). The name stuck, which is visible in several files in the repository.

Use example

Here is an exemplary (minimalistic) code for constructing a graph neural network of type 1:

import numpy as np
import tensorflow as tf

import graph_utils as graph_utils
import graph_neural_networks as graph_nn

adj_matrix = np.genfromtxt("macierz_sasiedztwa.txt")

nn_input = tf.placeholder(dtype=tf.float32, shape=[None, 21])
nn_output = graph_nn.transfer_matrix_neural_net(nn_input, 3, 4, tf.nn.tanh, adj_matrix, verbose=True)

References

• van der Maaten, L. and Hinton, G. (2008) Visualizing Data using t-SNE, Journal of Machine Learning Research, vol. 9, pp. 2579--2605
• Kipf, T. N. and Welling, M. (2017) Semi-Supervised Classification with Graph Convolutional Networks, 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings