mmao-opt 0.2.0

A metabolic multi-agent optimization framework for continuous, discrete, dynamic, and classification-oriented search.

mmao

MMAO is a metabolic multi-agent metaheuristic optimization framework for continuous optimization, combinatorial optimization, dynamic optimization, and classification-oriented search. It uses endogenous resource allocation, closed-loop adaptation, and parameter-light control to provide a reusable Python package for optimization research, algorithm development, and reproducible benchmarking. The repository includes benchmark loaders, command-line tools, example workflows, and research-friendly APIs for tasks such as TSP, multidimensional knapsack, feature selection, and classifier tuning.

The table below summarizes how MMAO differs, at the framework level, from several well-known heuristic and plant-inspired optimizers. The goal is not to rank these methods universally, but to clarify the design dimensions on which MMAO contributes a distinct closed-loop control logic.

Algorithm	Core Inspiration / Abstraction	Resource Allocation Logic	Population / Agent Lifecycle	Search Guidance	Adaptation Style	Cross-Domain Unification
MMAO / Metabolic Multi-Agent Optimizer	Metabolic closed-loop economics of survival and reinvestment	Endogenous private-public resource loop with earned, recycled, and reinvested budget	Dynamic branching, pruning, and respawn emerge from energy state	Energy-regulated sensing; symmetric probing in continuous spaces and structural sensing in discrete spaces	Unified endogenous adaptation of scale, role, effort, and redistribution	Yes; same control logic across continuous and discrete domains
PSO / Particle Swarm Optimization	Swarm motion and social learning	No endogenous resource economy	Usually fixed population	Velocity update using personal and global best	Usually external or heuristic parameter adaptation	Mainly continuous; discrete variants require redesign
DE / Differential Evolution	Differential variation in evolving populations	No endogenous resource accounting	Usually fixed population	Differential mutation, crossover, and greedy selection	Usually external parameter adaptation	Mainly continuous; discrete transfer is non-native
APOA / Artificial Plant Optimization Algorithm	Artificial plant growth and branching	No unified budget loop	Rule-based growth and branching	Plant-like branching rules	Mostly rule-based	Typically domain-specific
TSA / Tree-Seed Algorithm	Tree growth and seed dispersal	No endogenous reinvestment economy	Static individuals with seed generation	Seed dispersal around parent solutions	Mostly fixed probabilistic adaptation	Mainly continuous
IWO / Invasive Weed Optimization	Weed colonization and reproduction	No closed resource loop	Rule-driven expansion and reduction	Random dispersal with variance schedule	Schedule-based adaptation	Mainly continuous
RMO / Root Mass Optimization	Root morphology and branching geometry	No private-public budget closure	Geometry- or probability-driven branching	Root-like extension in search space	Morphology-driven adaptation	Mostly continuous or domain-specific
RGA / Root Growth Algorithm	Root growth and branching behavior	No endogenous budget economy	Branching governed by root-style rules	Root-direction and branching heuristics	Rule-based adaptation	Mostly continuous or domain-specific

What This Repository Provides

This repository is the unified software release for the MMAO research line. It consolidates the core ideas validated across the current MMAO papers into one installable Python package, with a shared API and lightweight command-line entry points.

The current package includes:

• Continuous optimization with a reusable MMAO controller
• Discrete optimization for TSP and multidimensional knapsack
• Dynamic optimization with environment-shift response
• Classification-oriented optimization for feature selection and classifier tuning
• Standard loaders for TSPLIB and OR-Library style benchmark data
• JSON result export, smoke tests, and a PyPI-ready package layout

Why This Repository Matters

The MMAO papers address different scientific questions, but external readers should not have to navigate several disconnected codebases. This repository is the common implementation surface that the papers can cite for:

• installation and quick verification
• unified interfaces across domains
• lightweight reproducibility
• standard data loading
• future benchmark expansion

Design Goal

This repository is intended to make MMAO easier to verify, test, and reuse. It is not just a dump of paper scripts. The package is organized around a common metabolic control logic:

• agents maintain private energy
• a communal resource pool redistributes search pressure
• role states shift continuously with success and scarcity
• expansion, contraction, and respawn are driven by the same internal loop

Installation

Install from PyPI with pip:

pip install mmao-opt

Install from PyPI with uv:

uv pip install mmao-opt

If you use uv as your project manager, you can also add the package as a dependency:

uv add mmao-opt

For local development with pip:

pip install -e .

For local development with uv:

uv pip install -e .

If you want TSPLIB loading support with pip:

pip install -e .[io]

If you want TSPLIB loading support with uv:

uv pip install -e ".[io]"

For development and testing with pip:

pip install -e .[dev]

For development and testing with uv:

uv pip install -e ".[dev]"

Quick Start

Run the packaged CLI:

mmao continuous --problem rastrigin --dimension 10 --iterations 80
mmao tsp --iterations 100
mmao dynamic --problem dyn-rastrigin --dimension 10 --iterations 80
mmao classification --dataset breast_cancer --classifier svm_rbf --iterations 12

To print a compact summary instead of the full history:

mmao continuous --problem sphere --dimension 5 --iterations 20 --summary-only

To save the full JSON result:

mmao continuous --problem sphere --dimension 5 --iterations 20 --output outputs/sphere.json

Standard data loaders are also available:

mmao tsp --tsplib /absolute/or/real/local/path/to/berlin52.tsp --summary-only
mmao knapsack --orlib /absolute/or/real/local/path/to/mknap2.txt --instance-index 0 --summary-only

The --tsplib and --orlib arguments expect real local file paths. If you just want a quick verification run after installation, use the bundled benchmarks below.

The package also bundles two standard benchmark files for direct verification after installation:

mmao tsp --benchmark berlin52 --summary-only
mmao knapsack --benchmark mknap2 --instance-index 0 --summary-only

If you cloned the GitHub repository, you can also run the local example script from the repository root:

cd mmao
python examples/quickstart.py

Python API

from mmao.api import optimize_continuous
from mmao.continuous import MMAOContinuousConfig, rastrigin_problem

problem = rastrigin_problem(dimension=10)
result = optimize_continuous(problem, MMAOContinuousConfig(iterations=80, seed=7))
print(result["best_fitness"])

For file-based benchmark loading:

from mmao import load_tsplib_problem, load_orlib_mkp_instance, optimize_tsp, optimize_knapsack

tsp_problem = load_tsplib_problem("berlin52.tsp")
mkp_problem = load_orlib_mkp_instance("mknap2.txt", index=0)

The short names "berlin52.tsp" and "mknap2.txt" resolve to the bundled examples shipped with the package. If you want to load your own files, pass a real local file path instead.

How to Extend MMAO

There are three practical ways to build on mmao-opt, depending on whether you want to reuse it as-is, adapt it to new problems, or modify the optimizer itself.

If you only need MMAO as a research tool, you can install the package and call the public API directly. The exported interfaces in src/mmao/api.py allow you to define your own problem instances and run MMAO without changing the package source. In practice, this is the easiest path for plugging MMAO into custom benchmark scripts, engineering workflows, or evaluation pipelines.

If you want to adapt or improve the implementation, clone the repository and use an editable install:

git clone https://github.com/wolfbrother/mmao.git
cd mmao
pip install -e .[dev]

Or with uv:

uv pip install -e ".[dev]"

The current package structure is intended to make domain-specific extensions straightforward:

• src/mmao/continuous/core.py for continuous optimization
• src/mmao/discrete/tsp.py for TSP
• src/mmao/discrete/knapsack.py for multidimensional knapsack
• src/mmao/dynamic/core.py for dynamic optimization
• src/mmao/classification/core.py for classification-oriented optimization
• src/mmao/io/ for benchmark and dataset loaders

Typical extension directions include:

• replacing or strengthening sensing and move operators
• refining branching, pruning, respawn, or reinvestment logic
• adding new benchmark adapters and data loaders
• introducing new problem families such as constrained, multi-objective, or scheduling optimization
• reducing manual hyperparameters by deriving more behaviors from the internal metabolic loop

If you plan to publish a derived variant, the most important design recommendation is to preserve the core MMAO identity: private energy, a communal resource pool, and closed-loop adaptation should remain the primary control mechanism. This makes it easier for others to understand the new method as part of the MMAO family rather than as a disconnected heuristic.

Package Structure

mmao/
  continuous/      continuous optimization
  discrete/        tsp and knapsack
  dynamic/         dynamic optimization
  classification/  feature selection and classifier tuning
  io/              optional loaders such as TSPLIB
  cli.py           command-line interface

Scope of This First Release

This first release focuses on a clean and usable public package interface. It already exposes the major MMAO problem families, but it is still a research-oriented release rather than a fully stabilized production library.

Near-term improvements will likely include:

• richer benchmark adapters
• experiment orchestration utilities
• result serialization helpers
• stronger dataset and problem loaders
• broader documentation for paper-level reproducibility

Testing

Run the smoke tests with:

pytest

The repository currently includes:

• API smoke tests
• CLI smoke tests
• data-loader tests
• source and wheel build validation

For public trust and regression control, CI is defined in .github/workflows/ci.yml.

Citation

Please use CITATION.cff for software citation metadata.

Recommended practice:

• cite the repository or mmao-opt package for software availability and reproducibility
• cite the foundational MMAO paper for the core optimizer
• cite the relevant derivative paper for domain-specific discussion such as dynamic optimization or classification-oriented search

Additional orientation documents:

• docs/paper-coverage.md
• docs/reproducibility.md