DPSO: Divergence-Guided Particle Swarm Optimization

A JAX-based implementation of DPSO, a PSO variant that adds a divergence-based repulsion term to prevent premature convergence on multimodal landscapes.

Quick start

uv sync
uv run python main.py                          # run experiments + generate artifacts
uv run python main.py --from-raw 2026-03-15_143022  # regenerate plots/tables from saved run

How DPSO works

DPSO augments the standard PSO velocity update with a modulation term:

$$ v_{mod} = c_3 * r_3 * \kappa(p_i, g) * \hat{d} $$

where $\kappa$ is a Gaussian similarity kernel between a particle's personal best and the global best, and $\hat{d}$ points away from the global best. Particles whose personal bests have converged near the global best receive a repulsive push, preserving exploration. Setting $c_3 = 0$ recovers standard PSO.

License

MIT