title	author
Bout duration distributions in animals	Pranav Minasandra

Analysing behavioural dynamics in the wild.

Details

The analyses in this code forms the basis of our pre-print, 'Behavioral sequences across multiple animal species in the wild share common structural features'.

People who worked actively on this code

Pranav Minasandra

People who performed a code review

Katrina Brock and Ariana Strandburg-Peshkin

People whose contributions were necessary for this project to get going

Ariana Strandburg-Peshkin, Emily Grout, Katrina Brock, Meg Crofoot, Vlad Demartsev, Andy Gersick, Ben Hirsch, Kay Holekamp, Lily Johnson-Ulrich, Amlan Nayak, Josue Ortega, Marie Roch, Eli Strauss, Marta Manser, Frants Jensen, Baptiste Averly, and many others

Overview

This project ties together results from behavioural classifiers built using hyenas, meerkats, coatis. Here, I find long timescale patterns in behavioural dynamics for all classified behaviours for each individual of each species. This project stems from my serendipitous discovery of heavy-tailed bout duration distributions in spotted hyenas in 2019.

Heavy-tailed distributions of bout durations could imply that self-reinforcement plays a role in behavioural dynamics at the fine scale: such distributions have decreasing associated hazard rates, which means that the longer the animal is in a behavioural state, the less likely it becomes to exit that behaviour in the next instant. This discovery implies that wildly different mammals have decreasing hazard rates for all behavioural states. We also show that all bout duration distributions are near power-law or truncated power-law types. We also find the self-dependence of behavioural time-series by quantifying rigorously mutual information between behaviour now and behaviour in the future, and ask how this mutual information decays with time. We show that the memory of a time-series of behaviour decays as a power-law up to a point (around 1000-3000 s), after which it changes to a more typical exponential decay. We show this in many different ways (check out our pre-print above!)

We use the module powerlaw for distribution fitting. Hazard functions are estimated by our code based on the definition of a hazard function. Mutual information is computed using sklearn, and a custom implementation of a Bialek finite-size correction is used to extrapolate its true value. Data for this project comes from the Communication and Coordination Across Scales interdisciplinary group.

Dependencies and prerequisites

This software has been written in python 3.10 on and for a Linux operating system. You will not need expensive supercomputers to run this code, it should work on any personal computer (tested on i7-11th Gen, 16G RAM). However, it will take very long. A more powerful computer will help to get these analyses replicated in a reasonable time. At the end there are tips for using this code on a less powerful computer. I have also tried to make this as OS-agnostic and IDE-agnostic as possible, so you should be able to run this on any computer directly. However, this code was written and run on Linux systems using the Bash shell, and I have only tested it in this configuration.

Below are details about how to install and run this software

Pre-requisite software

The following packages have to be installed separately:

• matplotlib
• numpy
• pandas
• powerlaw
• scikit-learn (for metrics)
• scipy
• nolds (for DFA)

Use pip to install these softwares, or any other package manager of your choice.

Installation and setup

NOTE: On Linux and (possibly Mac), several below steps are automated by running the following command.:

curl -sSf https://raw.githubusercontent.com/pminasandra/bout-duration-distributions/master/setup.sh | bash

It might fail if your version of pip is old; so try updating that if there is a pip related error. If you have run the above command, skip straight to step 5.

1. create a project directory at a location of your choice and enter it

mkdir /path/to/your/project
cd /path/to/your/project

2. Download the contents of this repository using

git clone https://github.com/pminasandra/bout-duration-distributions code

3. Also create the Data and Figures directories

mkdir Data
mkdir Figures

4. In the code/ directory, create a file called 'cwd.txt' that, on the very first line, has the content /path/to/your/project

You can do this in linux-like command lines like this:

echo $PWD > code/cwd.txt

5. After this, copy any behaviour sequence data folders into the Data/ folder.

Usage

Run all indicated python scripts using a terminal, with the command python3 <script_name>.py

Analyses are to be done as follows:

• Running python3 replicates.py generates Markovised pseudosequences to be used in the remaining analyses. This step will use a lot of storage space. Ensure that you have enough free space on your hard-disk.
• Running python3 code/pkgnametbd/fitting.py generates all bout duration distributions and generates tables containing AIC values.
• Running python3 code/pkgnametbd/survival.py creates plots with the hazard functions for all behaviours.
• Running python3 code/pkgnametbd/persistence.py performs DFA and mutual information decay analyses.
• Running python3 code/pkgnametbd/simulate.py runs all simulations mentioned in the paper and its appendices.

For academic colleagues, it is easy to re-work this code in your own analyses. I am also hoping to soon make a more friendly library to incorporate these analyses in your workflow. Most functions also come with helpful docstrings, and the overall code structure is modular and intuitive. If you are familiar with basic python, the only additional thing you need to know is about generators, a python object that is not commonly used, but speeds up work tremendously in our case. Useful classes are provided by simulations/agentpool.py and simulations/simulator.py, and generally helpful functions are found in boutparsing.py and fitting.py.

Usage on less powerful computers

With the implementation of Markovised pseudosequences, bootstrapping, and Bialek correction of mutual information, this code now demands a lot of computational power. However, good replicable code should be designed so that it can be run on any system for testing. This also keeps science equitable. Here are some tips for running this code on a local, non-super-computer. In config.py

• Decrease NUM_MARKOVISED_SEQUENCES. Something like 3 is a good number to attempt replication.
• Decrease NUM_BOOTSTRAP_REPS. Something like 10 is a good number.
• If you desire, set ADD_MARKOV and ADD_BOOTSTRAPPING to False. This will greatly speed up your code.
• Increase of decrease NUM_CORES based on how many processes you can run in parallel on your machine.