April 20, 2026
Craig Reynolds
cwr@red3d.com

This is experimental code, described in a draft paper EvoFlock: evolved inverse design of multi-agent motion, currently under review.

The abstract of the abstract is that multi-agent motion models (such as boids) can be hard to adjust. They have many (~15) parameters which interact in nonlinear ways.

This work takes an inverse design approach using optimization (via a genetic algorithm) according to an objective function (written by the user).

So far, I have treated this code as a private branch: it is often broken, or modified for debugging. If you are interested in building/using it, (a) thanks for your interest!, and (b) please send me email so I can help you get a clean version. If you do try, please let me know how you fare so I can make this code easier to use. For example, I think it has only been run on macOS.

Just a few hints:

• Build EvoFlock using either:
  • CMake on CMakeLists.txt
  • Xcode on evoflock.xcodeproj
• Dependencies (these should be formalized as git submodule dependancies):
  • EvoFlock uses an evolutionary computation engine, called LazyPredator, developed for an earlier project.
    • Its updated source code is in a subdirectory: EvoFlock/LazyPredator/.
    • LazyPredator supports genetic programming (GP). A genetic algorithm (GA) can be considered a special case of GP.
  • The Open3D library, a modern layer on top of OpenGL, is used for graphics. You will need to install this separately.
• Run either from Xcode or with the evoflock executable from the command line.
  • The normal action is to start optimization run, finding a solution for the given objective. For a GA run this will be a set of scalar values in a ParameterSet object, or in the experimental GP mode, the result will be “source code” in a domain specific language.
  • Otherwise if EF::visualize_previous_results_mode is true will run flock simulations based on results from a previous optimization run.
  • Normally an Open3D window will open and logging will appear on the shell where evoflock was run.
    • A simple user interface is available in the Open3D window
    • The mouse can be used to adjust the “camera” (“point of view”).
    • A few single characters commands exist. For example type an "H" to the Open3D window to print a list of currently defined command. See below this paragraph. The “B” key command can be used during an optimization run to check on its progress. The optimization is paused, a flock simulation is run with the best solution found so far, then the optimization resumes.

        EvoFlock: list of single key commands.
        H  print this list of single key commands.
        G  toggle "graphics mode".
        C  cycle through camera aiming modes.
           [space] toggle simulation pause.
        O  cycle through predefined obstacle sets.
        1  single step mode (advance one simulation step, then pause).
        S  cycle selected boid through flock.
        A  toggle drawing of annotation lines, etc.
        R  reset camera to aligned view of whole scene.
        B  pause, run sim with best individual & graphics, then proceed.
        W  toggle "space-time worms".
    

• EvoFlock has several versions and modes of operation, all folded into a single code base using mode switches which are defined in EvoFlock.h. The progression over time was:
  • Use GA to find a ParameterSet for parametric flocking model according to three objectives: correct separation from nearest neighbor, obstacle avoidance, maintaining a target speed.
  • Add an objective to increase each boid's path curvature (EF::use_curvature_objective).
  • Switch between GA and GP (EF::usingGP(), EF::usingGA()) for ongoing experiments.
  • Switch between earlier “flocking with obstacle” version and experiments with murmurations (EF::murmuration_mode).
• Primary evolution and simulation parameters:
  • in EF::runOneFlockEvolution():
    • Evolutionary population: 300
    • Evolutionary steps: 30000 (for SSGA, essentially 100 “generations”)
  • in FlockParameters:
    • Boids per flock: 200
    • Steps per flock simulation: 500
• I hope to soon convert this messy personal repository to be useful to others…