gmpb 0.1.0

A Python implementation of the Generalized Moving Peaks Benchmark

gmpb

This Python implementation follows the Generalized Moving Peaks Benchmark introduced by Danial Yazdani, Michalis Mavrovouniotis, Changhe Li, Guoyu Chen, Wenjian Luo, Mohammad Nabi Omidvar, Juergen Branke, Shengxiang Yang, and Xin Yao.

gmpb is a small, reusable, MIT-licensed implementation of the fully non-separable GMPB formulation described in the paper “Competition on Dynamic Optimization Problems Generated by Generalized Moving Peaks Benchmark (GMPB)”. The core library intentionally does not depend on or derive its implementation from the original MATLAB/Octave codebase.

Quick Start

Requires Python 3.10+.

python -m venv .venv
source .venv/bin/activate
make dev
make test

import numpy as np

from gmpb import GMPB, GMPBConfig

config = GMPBConfig(d=5, m=10, change_frequency=5000)
problem = GMPB(config, seed=7)

value = problem.evaluate(np.zeros(config.d))
batch = problem.evaluate_batch(np.random.default_rng(0).normal(size=(32, config.d)))

Public API

The stable import surface for 0.1.x is:

from gmpb import GMPB, GMPBConfig
from gmpb.metrics import end_of_env_error, offline_error

Stepping Semantics

GMPB is a dynamic maximization benchmark. With the default auto_step=True:

• Each scalar or batch evaluation consumes fitness evaluations.
• The change_frequency-th evaluation is still scored in the current environment.
• The benchmark steps immediately after returning the value that consumed the last evaluation in that environment.
• evaluate_batch() splits work into chunks when a batch crosses environment boundaries, so each sub-block is scored under the correct environment.

If you want to control environment changes manually, set auto_step=False and call step_environment() directly.

Metrics

The benchmark does not compute the true optimum value f* for each environment. You supply those values externally when calculating:

• offline_error(f_star, best_so_far)
• end_of_env_error(f_star, best_end)

Minimization

GMPB is defined as a maximization problem. To use it in a minimization pipeline, negate the returned objective values.

Examples And Docs

Run the minimal example:

make run-example

Build the docs:

make docs

The HTML output is written to docs/_build/html/.

Upstream Compatibility

This repository can optionally validate its installed Python implementation against EDOLAB's Octave GMPB implementation as a test-only oracle. EDOLAB is not a runtime dependency and is never bundled into the package artifacts.

For a local compatibility run, clone EDOLAB into _edolab/, check out the pinned commit in tests/upstream/EDOLAB_SHA.txt, and run:

make test-upstream-compat