Code and data to reproduce results from "Tuning movement for sensing in an uncertain world" by Chen Chen, Todd D. Murphey, and Malcolm A. MacIver, Northwestern University, Evanston IL, USA (eLife, 2020).

Ergodic Information Harvesting (EIH) Video & Tutorial

• Video showing the fish, mammal, and insect behaviors that were compared to the trajectories generated by EIH, and a video showing how the algorithm works by way of applying it to control an underwater electrolocation robot.

• Interactive Jupyter notebook tutorial, click to view online:

  or use to run interactively through online Jupyter Notebook

Steps to reproduce the results shown in the EIH paper

All of the simulation code is written with Python 3. All of the figure plotting files are written in MATLAB (R2017a+). The code can be run on:

• A local computer, which is very easy to set up but the performance is ultimately limited by the number of locally accessible CPU cores
• Cloud computing virtual servers through any Infrastructure as a Service (IaaS) provider, e.g. Amazon Elastic Compute Cloud, Google Cloud Compute Engine, or academic HPCC (High Performance Computing Cluster) systems. Cloud computing is easy to setup and provides a way to scale up the total number of running threads (e.g. Google Cloud Compute Engine allows up to 96 CPU threads per instance). Our code's runtime environment and dependencies are fully containerized through Docker to minimize the effort needed for environment setup, and for easy scaling to run the code faster (you can run the code on a single CPU over a few days, or on many CPUS on the cloud in a few hours). Setting the code up to run on a cloud service for the first time is somewhat involved if you are not used to doing this. We have a set of screencasts to walk through doing this from scratch on Amazon Web Services that we are happy to share.

Note that this repository has included all of the published data, including all the simulations, to reproduce all figures in the paper. To reproduce our figures from the published data, rather than re-run all the simulations from scratch, simply jump to step 5 below.

Detailed Steps

To avoid possible dependency hell and minimize the effort of setting up the runtime environment we used for our results, we prebuilt a container image to be used for executing all the simulation code in Python using Docker. Here is an article explaining the utility of containers for reproducibility of research: Singularity: Scientific containers for mobility of compute. Note that this is only for reproducing simulations: for generation of the figures from the simulations, a local installation of MATLAB (not provided in the container) is still required.

1. Obtain code and data

To start, you will either be using the Zenodo archived repository, which includes both the data and the code; or, if you prefer, you will clone the most recent version of the EIH repository. To obtain the Zenodo archived repository, which is a zip of the release of the repository corresponding to the release of the publication by eLife, you will simply search "maciver chen" on the Zenodo website, and download the 32GB zip file of this repository.

To clone the most recent version of the EIH repository (which may have changes from the archived release to correct errors that are found), two tools are required:

• git - is used to pull all the non-data files from this repository. Go to git's official release page to download and install git. Then use the following command to clone this repository:

git clone --depth=1 https://github.com/MacIver-Lab/Ergodic-Information-Harvesting

• git-lfs - is used to pull all the published data (required to reproduce the results). Go to git-lfs's official release page to download and install. Then run the following command inside the root directory of the cloned EIH repo ./Ergodic-Information-Harvesting/

cd Ergodic-Information-Harvesting
git lfs install
git lfs pull

If you succeeded, you should see files being downloaded by git-lfs. Once it is setup, should you decide to delete the files and start again, you should only need to do the git clone step.

2. Install Docker and Pull the EIH Container Image

• Install Docker by following the official documentation: https://docs.docker.com/engine/install/

• (Optionally) For Linux or Linux-based HPCC/cloud computing environments, please additionally follow the post-installation setup steps for Linux to allow running docker without sudo. If you don't want or unable to follow this step, you will need to make sure to run docker commands with sudo docker rather than docker.

3. Invoke Shell in the EIH Container Image

The container image is a fully self-contained Linux OS image with Python 3 dependencies setup for generating the EIH simulations developed for the study. We will invoke the command line tool inside of the EIH container image to interact with the resources inside the container and start the simulations.

docker run -it -v [absolute path to EIH folder]:/EIH maciverlabnu/ergodic-information-harvesting

replace the [absolute path to EIH folder] part with the absolute path to your local EIH repository folder, e.g. C:/Ergodic-Information-Harvesting (remember to replace \ with / when in Windows) or ~/Ergodic-Information-Harvesting.

If you are already inside the Ergodic-Information-Harvesting folder, you can simply do

docker run -it -v $PWD:/EIH maciverlabnu/ergodic-information-harvesting

4. Start Reproducing Simulation

We used Cython to accelerate the simulation which requires compiling some of the code before running the simulation. Compile the accelerated code by calling the following command (this only needs to be done once):

# Command to run inside the container
cd /EIH
chmod +x ./BuildCython.sh
./BuildCython.sh

You are now all set for the environment setup. You can start reproducing all the simulation results by running the main simulation code:

cd /EIH/SimulationCode/
python3 RunAllSims.py

By default, RunAllSims.py will check the number of available CPU threads and automally run parallel simulation jobs with the maximum number of threads possible. Nonetheless, the number of threads can be manually specified by passing the desired parallel thread count argument to it, for example

python3 RunAllSims.py 20

will run 20 threads in parallel.

NOTE: The simulation will take a long time to finish. Depending on your operating system, you may need to prevent your system from going to sleep. This is necessary with MacOS. With MacOS: Open a terminal, and type caffeinate and hit return. Your system will be prevented from sleeping until you hit Control-C.

Once all the simulation jobs are done, exit the Singularity shell environment by calling the exit command.

5. Reproduce Figure Results

The figure generation code is written in MATLAB and MATLAB R2017a or a more recent version is required. To start, open the makeFigurePanels.m code in MATLAB under the Production-Figure-Code folder. To reproduce figure 2, for example, use the following procedure:

• Launch Ergodic-Information-Harvesting/Production-Figure-Code/makeFigurePanels.m using MATLAB. Note that the code has been tested with MATLAB R2017a and R2018a.
• Specify input parameters
  • Set targetFig = 'fig2' to select figure 2 as the target
  • Set USE_PUBLISHED_DATASET = 1 to use the published dataset included in the repository. Alternatively, if the local simulation jobs are completed, use of USE_PUBLISHED_DATASET = 0 will force the code to use reproduced data located at Ergodic-Information-Harvesting/SimulationCode/SimData/
• Run the MATLAB code

You should see a new MATLAB figure containing Figure 2 panels. PDF(s) will be saved under Ergodic-Information-Harvesting/Production-Figure-Code/FigureOutput/fig2/.

To reproduce all the figures, follow the same steps, but set targetFig = 'all'

Benchmark Running Time

• Benchmark on AWS: ~4.03 hours on AWS EC2 c5a.24xlarge instance (Ubuntu 18.04 LTS 64-bit, AMD EPYC 7R32 with boost frequency up-to 3.3GHz, 48 Core/96 Threads available under HVM), ~22.17 hours on AWS EC2 c5a.4xlarge instance (Ubuntu 18.04 LTS 64-bit, AMD EPYC 7R32 with boost frequency up-to 3.3GHz, 8 Core/16 Threads available under HVM). Benchmark on macOS: ~100 hours on a 2015 iMac (macOS 10.15.6, 64-bit 3.3 GHz Quad-Core Intel Core i5, 4 Core/8 Threads).

Additional Note for Linux and MacOS Users

Prevent System from Sleeping During Simulation

To prevent MacOS from sleeping in these instances, use caffeinate at a Terminal window running simulation jobs.