An algorithm called Mycelial Search, or Myco, reached a mean final error of zero, with a standard deviation of zero, on the F1 test function at both dimensions examined. At D = 10, it tied with jSO, L-SHADE and JADE. At D = 20, it again recorded zero mean final error on F1 and the lowest reported mean on F2, where the mean final error was 29.9 and the standard deviation was 22.7. These were the preprint’s clearest reported results, while performance varied across functions.
The test was a computational comparison using the CEC 2022 functions F1 through F11 at D = 10 and D = 20, where D is the number of dimensions in a problem. The study ran each algorithm-function pairing 30 times and reported the average final error alongside its standard deviation, a measure of spread across runs. Myco used 30 tips per run, while the comparison algorithms used 50 individuals. The evaluation limits were 200,000 function evaluations at D = 10 and 1,000,000 at D = 20, with budgets described as 20,000D and 50,000D. The comparison included jSO, SAP-DE, L-SHADE, JADE, CLPSO, ABC, GA, GWO, WOA, SMA and MGO.
A search built from local links
Myco represents candidate solutions as active tips and retained anchors in a spatial graph that is rebuilt at every iteration. Tips receive locally weighted flow from neighbours, while anchors preserve high-quality locations. The algorithm can inject those anchors back into the search, supplementing the movement driven by the tips. The result is a system that continually redraws local connections while keeping a record of strong positions.
A Louvain partition, a way of dividing a graph into communities, separates exchange within a community from exchange across communities. Myco also uses adaptive cord plasticity: tip-to-tip edges that align with the flow can be reinforced or allowed to decay. The authors describe these as operating at different scales. Community detection sets the range of information exchange at the partition level, while plasticity works at the edge level to reinforce or weaken flow-aligned connections.
The result depended on the landscape
The strongest and weakest cases appeared to depend on the shape of the test landscape, according to the authors. Myco could remain competitive when local flow identified a productive direction, but performance weakened when useful information was spread across disconnected or competing regions. The distinction matters: the method appears better suited when nearby signals point toward progress than when promising clues are separated or pull the search in different directions.
An ablation, meaning a test in which one part of a method is removed or changed, produced a mixed picture. In the cord-plasticity comparison, the largest reported mean-error differences at D = 10 were on F4 and F6. At D = 20, the Plasticity variant had lower mean error on F1, F3, F4, F5, F7, F8, F9 and F11. The paper says the benefit was not universal.
The Louvain version remained preferable to Plasticity on F2, F6 and F10 in some landscapes. The paper interprets those cases as suggesting that letting link strength decay may weaken alternative directions before their value becomes clear. That is a reminder that preserving a locally promising route and allowing other routes to survive are not always the same thing.
A warning from the alternative graph
A separate sensitivity test replaced the Louvain community backend with Greedy Modularity. Across all reported functions and dimensions, Greedy Modularity produced one raw community and one retained community. That removed the distinction between within- and cross-community weighting, so all present edges received the same community factor. The finding describes this reported backend configuration, not a general property of Greedy Modularity.
In selected Greedy diagnostics, the mean-to-median ratio became very large. The diagnostic is written as : the Greek letter mu in the numerator denotes the mean final error, and the x with a tilde in the denominator denotes the median final error. The reported ratios were 7.05 × 10^4 for D = 10 F1, 5.01 × 10^9 for D = 10 F11, 9.86 for D = 20 F10 and 1.59 × 10^12 for D = 20 F11. The selected cases show why the paper reports both mean and median behaviour when comparing the backend.
A bounded claim
The document is an arXiv version 2 preprint dated 31 August 2026. The benchmark comparison also used different population sizes, with 30 Myco tips per run versus 50 individuals for the other algorithms, a caveat when comparing the results. The paper states that its Python implementation is publicly available at the GitHub URL it provides. Overall, the study presents Myco as competitive on selected functions, with its broader prospects tied to how effectively its graph carries useful information between competing parts of a search landscape.
Paper data and sources
Original title: Mycelial Search: A Graph-Structured Metaheuristic for Continuous Optimisation
Authors: Mohammad Mahdi Dehshibi
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text