Hyperparameter sensitivity

PSO has several hyperparameters (w, c1, c2, swarm size, topology, …) and their best values are problem-dependent. turboswarm.sweep runs the optimizer over the Cartesian product of the value lists you give it, so you can see which hyperparameters actually move the solution quality.

Because PSO is stochastic, each combination can be repeated over several seeds and aggregated (mean / std / min / max of the best value).

Example

import turboswarm as pso

sweep = pso.sweep(
    "rastrigin", bounds=(-5.12, 5.12), dim=2,
    grid={"w": [0.4, 0.7, 0.9], "c1": [1.0, 2.0]},   # 3 x 2 = 6 combinations
    seeds=5,                                          # repeat each one 5 times
    n_particles=30, max_iter=100,                     # fixed for every run
)

print(sweep.best())          # combination with the lowest mean best value
for rec in sweep:
    print(rec["w"], rec["c1"], rec["mean"], rec["std"])

Each record is a dict with the swept hyperparameters plus mean, std (population), min, max, the raw values list and n (number of seeds).

Result object

sweep returns a SweepResult that is iterable and indexable over its records:

Member	Meaning
len(result)	number of combinations
result.best(metric="mean")	record with the lowest value of metric
result.to_dataframe()	the records as a pandas DataFrame (pandas imported lazily — not a hard dependency)

What you can sweep

The keys of grid are any keyword accepted by minimize — e.g. w, c1, c2, n_particles, max_iter, velocity, topology, v_max. Anything you do not sweep is passed as a fixed keyword argument and forwarded to every run. A key cannot be both swept and fixed, and repetitions go through seeds= (do not pass a fixed seed). History recording is turned off by default for speed.

Visualizing

viz.plot_sensitivity draws a line for one swept hyperparameter (with error bars from the spread) or a heatmap for two:

import matplotlib.pyplot as plt

pso.viz.plot_sensitivity(sweep, x="w")            # 1D line
pso.viz.plot_sensitivity(sweep, x="w", y="c1")    # 2D heatmap
plt.show()

When the sweep varies more hyperparameters than the ones plotted, the points are marginalized (averaged) over the rest.