GNBG-II

Generalized Numerical Benchmark Generator - II: A benchmark suite consisting of 24 test problems for the 2025 CEC and GECCO competitions.

This repository contains a GPU-accelerated implementation using Rust and WebGPU, with Python bindings. Performance measurements indicate 6.8x to 12.1x speedup over CPU implementations while maintaining numerical consistency.

Requirements

• Rust (stable)
• Python 3.11+
• Maturin

Installation

git clone https://github.com/minkymorgan/GNBG-II.git
cd GNBG-II
pip install maturin
maturin develop --features python

Usage

import gnbg_gpu
import numpy as np

# Problem F1
gnbg = gnbg_gpu.GNBGGpu(1, use_gpu=True)

# Single evaluation
solution = np.random.uniform(-100, 100, gnbg.dimension)
fitness = gnbg.fitness_single(solution)

# Batch evaluation
solutions = np.random.uniform(-100, 100, (100, gnbg.dimension))
fitness_batch = gnbg.fitness(solutions)

API

GNBGGpu(problem_index, use_gpu=True)

Parameters:

• problem_index: Problem number (1-24)
• use_gpu: Enable GPU acceleration

Methods:

• fitness(X): Batch evaluation
• fitness_single(x): Single evaluation
• reset(): Reset counters

GNBG(problem_index, use_gpu=True)

Python wrapper maintaining compatibility with original interface.

GNBG Function Reference (F1-F24 to GF Mapping)

The GNBG-II suite contains 24 carefully designed optimization problems. For multi-objective extensions, we use GF (GPU Functions) terminology:

Unimodal Functions (GF1-GF6)

GF	GNBG	Characteristics	Optimum	Max Evals
GF1	F1	Well-conditioned, separable	-1081.984	500K
GF2	F2	Moderately conditioned	-104.865	500K
GF3	F3	Ill-conditioned	-51.479	500K
GF4	F4	Non-separable, rotated	-930.063	500K
GF5	F5	Non-separable, shifted	-525.477	500K
GF6	F6	Non-separable, rotated+shifted	-337.509	500K

Single-Component Multimodal (GF7-GF15)

GF	GNBG	Characteristics	Optimum	Max Evals
GF7	F7	Many local optima	-874.478	500K
GF8	F8	Deceptive landscape	-318.801	500K
GF9	F9	Highly multimodal	-1030.301	500K
GF10	F10	Rugged surface	-884.736	500K
GF11	F11	Asymmetric basins	-843.147	500K
GF12	F12	Variable interactions	-401.147	500K
GF13	F13	Non-linear basins	-159.612	500K
GF14	F14	Conditioning+multimodal	-189.941	500K
GF15	F15	Complex landscape	-624.056	500K

Multi-Component Multimodal (GF16-GF24)

GF	GNBG	Components	Characteristics	Optimum	Max Evals
GF16	F16	5	Hybrid composition	-4208.486	1M
GF17	F17	5	Rotated composition	-4381.528	1M
GF18	F18	5	Non-separable hybrid	-4046.060	1M
GF19	F19	5	Complex interactions	-966.475	1M
GF20	F20	2	Dual-component	-1000.000	1M
GF21	F21	5	Advanced hybrid	-992.221	1M
GF22	F22	5	Extreme conditioning	-99.363	1M
GF23	F23	5	Maximum complexity	-99.388	1M
GF24	F24	7	Ultimate challenge	-99.917	1M

Running Specific Functions

To run any of the 24 GNBG functions in experiments:

# Single-objective optimization
gnbg = gnbg_gpu.GNBGGpu(problem_index=5)  # For F5/GF5
solutions = np.random.uniform(-100, 100, (100, 30))
fitness_values = gnbg.fitness(solutions)

# Multi-objective GNBG-MO (experiments/)
from gnbg_gpu.multi_objective import GNBGMultiObjectiveProblem
problem = GNBGMultiObjectiveProblem.from_gnbg_function(
    gnbg_function=5,  # F5/GF5
    n_obj=10,         # 10 objectives
    n_var=30          # 30 variables
)

Note: All functions use 30-dimensional search space [-100, 100]^30 with acceptance threshold 1e-08.

Performance

Objectives	GPU (eval/s)	CPU (eval/s)	Speedup
5	206,558	30,355	6.8x
10	408,057	60,786	6.7x
20	989,193	122,111	8.1x
30	2,196,089	181,536	12.1x

License

GNU General Public License v3.0

Original GNBG-II Authors:

• Danial Yazdani (danial.yazdani@gmail.com) - Copyright (c) 2023
• Vladimir Stanovov (vladimirstanovov@yandex.ru) - C++ Implementation (2024)

GPU Acceleration:

• Andrew Morgan (minkymorgan@gmail.com) - Rust GPU Implementation (2025)