Get the data from the paper

You can download it automatically using the included script download_data.sh.

This repository works with swarmulator branch SI_framework: https://github.com/coppolam/SI_framework

Note, the script download_simulator.sh will also download the appropriate swarmulator branch in zip format.

Install Python required packages

sudo pip3 install -r requirements.txt

Main scripts

These are all the "main" scripts that have been used to generate/process the data.

• main_generate_data_training.py: generates the training data for a specified study case to then train the micro-macro link and extract the transition model.

• main_generate_data_benchmark.py: generates performance data for random policies to see how random policies perform.

• main_nn_training.py: learns and evaluates the micro-macro link model

• main_optimization.py: optimizes a policy and benchmarks the results

• main_verification.py:verifies an optimized policy according to the conditions in propositions 1 and 2. Gives the data seen in Table 2 if supplied with the supplied data (see download_data.sh script)

• main_evolve.py: uses the DEAP package to do a "standard" simulation based evolution for a specified study case. To have it work properly, the evolution.setup() function may have to be altered with good mutation/mating parameters (by default, mutation is binary which is not useful for this case)

Recreating the figures in the paper scripts

Use these scripts to create the figures used in the paper

Make all the figures in one go

• plot_all.py: generates all the plots in the paper, use with supplied data to recreate all figures exactly (see download_data.sh script in order to easily download the data)

Individual figure scripts (may require additional inputs, see plot_all.py)

• plot/plot_nn.py: recreates figures 4
• plot/plot_model.py: recreates figure 5
• main_nn_training.py: recreases figure 6 (but can also generate data!)
• plot/plot_benchmark.py: recreates figure 7
• plot/plot_logs.py: recreates figure 8
• main_verification.py: recreates figures 9 and 10 (and also verifies propositions)
• plot/plot_evolution.py: recreates figure 11
• plot/plot_onlinelearning.py: recreates figure 12

Packages

Classes

• verification.py: import and use the "get" function to get the conditions needed to run each study case properly

• simplenetwork.py: a higher level PyTorch class that sets up and runs a feedforward neural network. Change this class if you wish to change the parameters of a neural network

• pagerank_optimization.py: given a transition model and a set of desired observations, the "main" function in this class will optimize a policy and return the optimized policy according to Equation 6 the paper.

• evolution.py: a higher level DEAP API that sets up and runs the evolution. By default it has the settings to extract the desired observations.

• simulator.py: an API that can interface with swarmulator API at a higher level so that we can run and analyze all the simulations and logs easily

Tools

• swarmulator.py: This is a lower level Python API that can interface with the C++ simulator Swarmulator. The API can be used to set build and/or runtime parameters, build, read logs, start simulations, and more. A FIFO pipe is used by the API to read the fitness and the end of a run.
• graph.py: Graph functions built on top of the networkx package
• fileHandler.py: Functions to handle files, create folder, delete files, ect.
• matrixOperations.py: Functions for matrix handling, built on top of numpy
• prettyplot.py: Functions to help make prettier plots (for papers)
• fitness_functions.py: Re-implements fitness functions from swarmulator for post-processing, in case