Observing and Inferring A Collective

This repo contains the supporting code for the manuscript 'Observing a Collective to Infer the Characteristics of Agents'

The code is organized as two subdirectories: Simulation, which contains code (MATLAB) to generate simulated data, and Analysis, which contains code (Python) that performs classification analysis. The general workflow is as follows:

• Use simulation scripts to generate simulated data (MAT files).
• Use the functions in batch_processing.py script to generate summary data.
• Use the functions in figures.py to reproduce figures in the manuscript.

Simulation

Runner_OandI_delv_rho_Nr.m : Runs the 2D simulations for a variety of parameters: delv (which is the intrinsic speed, s0), Nr (which is the number ratio) and rho which is the packing density.

ABM_bidispese_delv_rho_Nr.m : Code for the Agent based model for circular agents in 2D periodic domain for a given set of parameters. agents_Expmemory_per2D.m : Contains the forces on the agents (self-propulsion, inter-agent short ranged interaction, brownian noise (turned off in the default))

RandomizationOfAgents_InitialConditions.m and agents_Expmemory_per2D_Randomization.m : These functions are used to create randomly packed arrangement of agents for the initial conditions to be used later in ABM_bidispese_delv_rho_Nr.m.

parameters.m and parameters_additional.m the required parameters for the ABM simulations.

NOTE: To reproduce the results in the paper (ArXiv link), run Runner_OandI_delv_rho_Nr.m. The system size N can be varied in the above m file. The parameters corresponding to the forces between the agents can be changed using the parameters.m file.

Analysis

Most of the heavy-lifting is done by the classes AgentDynamics (see agent_dynamics.py) and DataClassifier (see classify.py). AgentDynamics represents one simulation realization, while DataClassifier aggregates multiple realizations for a given set of parameters. See the methods of each class for more details, most of the methods are documented.

Once we have simulated data from the simulation scripts, batch_processing.py script contains functions to process and summarize classification results.

• cache_all_data and cache_all_data_parallel collects and summarizes simulation data (MAT files) into summary representations.
• compute_classification_metrics performs classification analysis (with both observers -- see paper for details) on the summary representations, and saves the confusion matrices.
• The functions in figures.py uses the confusion matrices and summary representations to generate figures from the manuscript.