---
kind: concept
slug: membrane-computing
title: Membrane computing
summary: >-
  Membrane computing models a cell as compartments that split and merge. In metaheuristic search, the population is split into membranes that search separately and are then merged. One author-reported card so far: it beats its base optimizer on a benchmark, at about 1.7 times the runtime, and its ablations lose to the full method.
tags: [optimization, membrane-computing, metaheuristic, p-system]
papers: ["2609.12756"]
related: [metaheuristic-optimization, particle-swarm-optimization, image-thresholding]
status: draft
updated_at: 2026-10-19
---

## Overview

Membrane computing is a model of computation inspired by cells. A P-system
holds objects inside compartments, called membranes, and applies rules that
split a membrane into several or merge them back [arxiv:2609.12756 §2.2]. The
card that covers this slug uses that idea as a population-management step
inside a nature-inspired optimizer, not as a model of biology.

The invariant is that the population of candidate solutions is always
recombined into one set after the membranes have searched separately. The
split is driven by a diversity value, not by a fixed schedule. Per §3.1 and
Algorithm 1, the population is split when that value is at or above a
threshold and stays whole below it [arxiv:2609.12756 §3.1]
[arxiv:2609.12756 alg. 1]. The paper's plain-language summary describes the
trigger the other way round, so this direction should be checked against the
paper before implementation.

This page rests on a single card. Every claim below is author-reported and
comes from one publication. No second paper in the wiki uses membrane
computing yet.

## Variants

### Membrane-split population in an artificial protozoa optimizer

The base method is the artificial protozoa optimizer. Its candidate solutions
forage, go dormant and reproduce [arxiv:2609.12756 §2.1]. The variant adds two
changes. A membrane step runs first. It measures a population diversity value,
mVOL, which is the population's spread divided by the search-box size, raised
to the power 1/dim [arxiv:2609.12756 eq. 15] [arxiv:2609.12756 eq. 16]
[arxiv:2609.12756 eq. 17]. If mVOL is below a threshold ε, the population
stays in one membrane. Otherwise it is split into m membranes, each updated
on its own, and the results are merged back [arxiv:2609.12756 §3.1]. The
number of membranes is drawn each iteration between 2 and 4 when m_max is 4
[arxiv:2609.12756 eq. 18].

The second change is in the guide selection. Instead of the usual guide, a
roulette wheel picks a candidate with probability proportional to a score that
mixes its normalised fitness and its normalised distance to the best candidate
[arxiv:2609.12756 §3.2] [arxiv:2609.12756 eq. 19]. The authors name this
roulette fitness-distance balance (RFDB) rule [arxiv:2609.12756 §3.2]. Card:
[MAAPO: an innovative membrane algorithm based on artificial protozoa optimizer for multilevel threshold image segmentation](/papers/2609.12756).

Reported results, author-reported, on CEC2017 at 10 dimensions with 51 runs
per function [arxiv:2609.12756 §4.1]:

- The variant has the best average Friedman rank over 13 algorithms, 3.35,
  against 3.37 for MAQUATRE and 3.55 for plain APO [arxiv:2609.12756 §4.5]. The
  margin over MAQUATRE is within 0.02 [arxiv:2609.12756 §4.5].
- Against plain APO, the variant wins on 8 of 30 functions and loses on none
  [arxiv:2609.12756 tab. 5].
- The full variant beats the membrane-only version and the RFDB-only version
  on 5 functions each, and loses on none [arxiv:2609.12756 tab. 5]. The paper
  therefore reports that the full combination beats each part on this suite.
- The average runtime is 1.38 s per run, against 0.81 s for plain APO
  [arxiv:2609.12756 tab. 7].

The same paper applies the variant to multilevel thresholding of three RGB
images, using Otsu and Kapur objectives. The author-reported ranks are mixed:
on Kapur it has the best average rank on all three metrics, while on Otsu it
is not best on PSNR [arxiv:2609.12756 tab. 11] [arxiv:2609.12756 tab. 12]. See
[image-thresholding](/concepts/image-thresholding).

## When to use which

There is only one variant, so this table compares the membrane approach with
the plain method it extends and states what the card does and does not
support.

| Situation | Use | Why |
|---|---|---|
| Continuous, box-bounded search where the base optimizer converges too early | Membrane split plus RFDB guide | The paper reports wins over plain APO on CEC2017 at 10-D [arxiv:2609.12756 tab. 5] |
| Tight time budget per run | Plain base optimizer | The variant takes about 1.7 times as long as plain APO on the same suite [arxiv:2609.12756 tab. 7] |
| Dimension other than 10 | Not decided by this card | CEC2017 is run only at 10-D, and the thresholding runs use 4 to 7 dimensions [arxiv:2609.12756 §4.1] [arxiv:2609.12756 §5.3] |
| Need for a top rank on Otsu PSNR | Another method, or test both | Author-reported average PSNR rank is 5.00 for the variant, behind five of the seven other methods in that table [arxiv:2609.12756 tab. 11] |
| Replicating the method exactly | Read the pitfalls before coding | No official code was found, and several rules are not stated [arxiv:2609.12756 §3] |

The paper's own parameter sweeps suggest a fixed ε near 0.0001 and m_max = 4
gave the best average ranks in its tests [arxiv:2609.12756 tab. 3]
[arxiv:2609.12756 tab. 4]. Those values are tuned on the same benchmark, so
they are not general advice.

## Implementation sketch

The loop below is the paper's Algorithm 1, condensed. The card does not
state how operators are chosen per solution, how the membranes are sized, or
which "best" the distance uses, so each of these is a choice the implementer
must make and document [arxiv:2609.12756 §3] [arxiv:2609.12756 alg. 1].

```pseudocode
# Random initial population of ps solutions, then evaluate  [§3]
population = sample_uniform(ps, dim, lower, upper)
fitness = evaluate(population)
while evaluations < budget:
    # Diversity value from the spread of the population  [eq. 15]
    mvol = compute_mvol(population, lower, upper)
    if mvol < eps:
        # Stay in one membrane, update all solutions  [alg. 1 lines 4-6]
        membranes = [population]
    else:
        # Split at random into m membranes  [eq. 18] [alg. 1 line 8]
        m = m_min + ceil((m_max - m_min) * rand())
        membranes = random_split(population, m)
    for each membrane in membranes:
        # Guide chosen by roulette over FDB scores  [eq. 19] [§3.2]
        for each solution x_i in membrane:
            x_i = apo_update(x_i, membrane)  # [eq. 21]
    # Merge membranes back into one population  [alg. 1 line 12]
    population = concatenate(membranes)
    fitness = evaluate(population)
    best = best_of(population, fitness)  # [alg. 1 line 14]
return best
```

Implementation notes from the card:

- The paper computes mVOL from products over dimensions. Working in log space
  avoids overflow and underflow at higher dimensions. This is our suggestion,
  not the paper's [arxiv:2609.12756 eq. 16] [arxiv:2609.12756 eq. 17].
- The score in eq. 19 is not stated to be negated for minimisation, and the
  roulette needs a non-negative score [arxiv:2609.12756 §3.2].

## Contradictions

None known. Only one card covers this slug, so there are no two papers to
compare. Within the paper, the ablation claim (membrane alone and RFDB alone
each lose to the full variant) and the Otsu PSNR result (the variant is not
best) are both author-reported and are not in conflict with each other.

## Papers

- [MAAPO: an innovative membrane algorithm based on artificial protozoa optimizer for multilevel threshold image segmentation](/papers/2609.12756): introduces a membrane split-and-merge step, driven by a diversity value, and an RFDB guide choice to the artificial protozoa optimizer.