Official Paper Repo for "Discrete Self-Adaptation in Competitive Coevolution for Constrained Hardware"

DOI: https://doi.org/10.1145/3795095.3805079

Abstract:

Self-adaptive competitive coevolutionary algorithms (CCEAs) typically rely on real-valued evolutionary parameters such as mutation rate. Yet emerging deployment targets, including specialized and resource-constrained hardware, often require integer-only or low-precision arithmetic which raises questions about how self-adaptation behaves under discrete constraints.

We investigate integer-only self-adaptive CCEAs, focusing on mutation rate and a Step Distribution Threshold (SDT) in an attacker–defender game, with results validated across Python and specialized-hardware implementations. Integer-valued parameters exhibit stable and effective adaptation, achieving convergence and performance comparable to continuous methods. Joint adaptation of mutation rate and SDT outperforms partially adaptive or fixed setups, maintaining a robust balance between exploration and exploitation across budget regimes.

We further show that integer-only self-adaptation responds reliably to environmental changes and extended evolutionary runs, and that these dynamics transfer to a pursuit–evasion style game. These findings establish discrete self-adaptation as a viable, generalizable mechanism for competitive coevolution on constrained hardware platforms.

Requirements

• Python 3.12 or newer
• Dependencies: Install the required packages using:

  pip install -r requirements.txt

• Git LFS: To be able to get the S1 data (paper_runs.zip <- 1.5 GB), you must install Git LFS (before cloning) using:

  git lfs install

• Storage: 10 GB of disk storage for the experimental data.

Running the Code

1. Unzip paper_runs.zip into the project root. This contains our experimental S1 data for the plots. The unzipped data requires approximately 7 GB of available disk space.

2. To generate the necessary Python data for the plots, run the experiment script using: python pythonCode/experiments.py. This will take a while to run.

3. Once the experiment script finishes, you can run all cells in paper_plots.ipynb to generate the paper plots. All figures will be exported to paper_runs/plots/. The final code block is computationally intensive and will take a few minutes to finish executing.