Mycelial Search: A Graph-Structured Metaheuristic for Continuous Optimisation

AI summary

Overview

Research area: Computational intelligence / nature-inspired metaheuristics for continuous numerical optimisation, specifically a graph-structured search algorithm inspired by fungal mycelial networks.

Technical level: Intermediate. The biological motivation and overall search concept are accessible, but the method combines Louvain community detection, conductance/flow dynamics, and gradient-based injection, so some familiarity with graph algorithms and population-based optimisation helps.

Scope: The paper introduces Mycelial Search (Myco), describes its graph-based mechanisms in detail, and benchmarks it against eleven established optimisers on the CEC 2022 single-objective bound-constrained suite at dimensions D=10 and D=20.

What This Paper Is About

Most population-based optimisers share information weakly — typically through a single global best, sampled exemplars, or temporary population differences — so they rarely keep a spatial interaction structure that changes as the search progresses. This can waste locally useful information, where a candidate would benefit from a nearby high-quality position without being pulled toward the same population-wide reference. Myco addresses this by treating candidate solutions as nodes in an evolving weighted graph, where local connectivity, community structure, and adaptive edge strengths determine what information flows where.

Key Contributions

1. A graph-structured metaheuristic for continuous optimisation (Myco). Candidate solutions are "active tips" plus a bounded set of "anchors" (historically favourable positions, capped at K_max), connected by edges induced by a fusion radius r_fuse = 0.1 ||u − l||_2. Unlike prior graph-oriented methods built for problems whose solution already is a network, here the evolving network becomes the search mechanism itself.

2. Community-weighted flow via Louvain partitioning. A Louvain partition (seed fixed at 42 for all reported experiments) distinguishes within-community (λ_intra) from cross-community (λ_inter) information exchange. Communities are filtered by a minimum node count (n_min) and induced subgraph density (ρ ≥ ρ_min); rejected nodes become singletons while keeping all their edges and contributing equally to flow.

3. Adaptive cord plasticity on tip-to-tip edges. Each undirected tip pair shares a single conductance scalar. Edges whose cosine alignment with the local flow exceeds τ_align are reinforced as g̃_ij = min(g_max, g_ij(1 + η s_ij)), then decay with lower clipping, g⁺_ij = max(g_min, (1 − δ) g̃_ij). Anchor-incident edges carry fixed conductance and are excluded from plasticity so anchors act as stable landmarks.

4. Ridge-Oriented Injection (ROI) as an auxiliary mechanism, plus a complexity analysis. ROI samples two anchors, computes a relative fitness gap γ, and — when γ > τ_roi — builds a spawn from the anchor midpoint using a central finite-difference gradient whose component orthogonal to the anchor axis is retained (x_roi = Π_[l,u](x_mid − α_roi p_⊥)). The paper derives a per-iteration cost of O(M log M + ED + ND) (M = N + K), reducing to O(N log N + ND) under sparse connectivity, and a total arithmetic cost of O(FE_max (log N + D)).

Main Findings

• Myco matched the best reported result on F1 at D = 10. Myco's mean final error on F1 at D = 10 was 0 ± 0, matching the boldface (lowest mean) values of jSO (0 ± 0), L-SHADE (0 ± 0) and JADE (0 ± 0).

• On F2–F7 at D = 10, the boldface lowest-mean error went to a differential-evolution comparator, not Myco. Myco's means were 6.14 ± 2.43 (F2, versus SAP-DE's 3.19 ± 2.98), 3.61 ± 0.418 (F3, versus JADE's 1.8 ± 0.373), 178 ± 737 (F4, versus JADE's 7.28 ± 3.38), 524 ± 617 (F5, versus L-SHADE and JADE at 0 ± 0), 92 ± 143 (F6, versus L-SHADE's 0.309 ± 0.218), and 1691 ± 4411 (F7, versus L-SHADE's 2.57 ± 13).

• The authors characterise the outcome as "competitive results on selected functions across both dimensions," not dominance. The abstract and introduction both frame the result this way, and the paper states that Myco "reaches competitive results on selected functions across both dimensions."

• Ablation: the two graph mechanisms play distinct roles. The ablation analysis shows that community structure regulates the range of graph-based information exchange, while cord plasticity controls the persistence of local directional influence.

• Overall interpretation. Graph-structured local interaction can support continuous optimisation, but its effectiveness depends on landscape structure and on how information is transferred across local search regions.

• Not reported in the available content: the full D = 20 results table (Table 4), the full D = 10 table beyond F7, the specific numerical outcomes of the ablation experiments, and the concrete values assigned to most hyperparameters (including η, δ, g_min, g_max, λ_intra, λ_inter, κ_min, κ_max, K_max, n_min, ρ_min, τ_align, τ_roi, α_roi, w, c_exp, c_flow).

Methodology in Plain English

The authors replace a flat population with a network. Each iteration, a fixed number of "tips" (Myco used 30 tips, while comparator implementations used their configured population of 50) sits in the search space alongside a small archive of "anchors" — the best positions seen so far, capped at K_max and distance-filtered (an anchor is only inserted if it is more than 10⁻⁶ away from every retained anchor). Tips that are close enough to each other become connected, and edge weight depends on both a stored conductance and the physical distance between nodes.

The network is then partitioned into communities using the Louvain method. Within-community and cross-community messages get different constant weights, and a sigmoid gate (whose steepness κ is annealed linearly from κ_max to κ_min over the run) scales each message by the potential difference between sender and receiver. This produces a flow vector at each tip: a weighted average of directions toward its neighbours, including anchors.

Next, edges between tips are reinforced or weakened based on whether they point along that flow, giving the graph a short-term directional memory. Tips then move by a velocity rule combining inertia, a pull toward the incumbent best (with a random element-wise multiplier), and the flow vector; isolated or near-isolated tips get a small uniform random perturbation with δ = 0.05 r_fuse. Optionally, ROI samples two anchors, and when their relative fitness gap is large enough, estimates a gradient by central finite differences at their midpoint (step size h = 10⁻⁵ max{|f(x_mid)|, 1}, costing 2D + 1 evaluations), removes the component along the anchor axis, and spawns a candidate in the transverse direction — replacing the worst tip only if the spawn is strictly better.

Evaluation used the CEC 2022 single-objective bound-constrained suite, functions F1–F11, at D = 10 (budget 20,000D = 200,000 evaluations) and D = 20 (budget 50,000D = 1,000,000 evaluations), with 30 independent runs per algorithm–function pair and search bounds of [−100, 100]. Random seeds were generated deterministically per function and run, and convergence was recorded at 16 checkpoints of the form ⌊D^(k/5−3) FE_max⌋ for k = 0, …, 15. Eleven comparators were grouped by family: differential evolution (jSO, SAP-DE, L-SHADE, JADE), swarm intelligence (CLPSO, ABC), genetic (GA), mammal-inspired (GWO, WOA), and decentralised growth (SMA, MGO). The Python implementation is released at github.com/dehshibi/Mycelia-Search.

Why This Matters

Impact on research. The paper tests whether explicitly modelling interaction structure — rather than only updating individuals — adds usable search information in continuous domains. It imports community detection and conductance-based reinforcement, both previously used mainly on network-centred problems, into a continuous metaheuristic, and reports those two mechanisms as separable in their effects. It also provides a complexity result (O(N log N + ND) per iteration under sparse connectivity) for graph-based search, which is a practical consideration for evaluating whether such methods scale.

Real-world applications (as motivated by the paper's own framing of continuous optimisation with interacting variables and multiple local optima):

• Constrained engineering design problems with interacting variables.
• Optimisation settings with multiple local optima, where maintaining alternative search directions matters.
• Problems with regions of sharply different search behaviour, where global guidance may be the wrong information to propagate.
• Network-flavoured optimisation tasks in general, given the algorithm's graph-native structure.

Industry relevance. The most direct industrial relevance is as a general-purpose continuous optimiser for expensive, non-convex, bound-constrained problems with a fixed evaluation budget — a common setting in engineering design and simulation-driven tuning. The paper does not report any industrial case study, only benchmark results.

Future Directions

• What drives the mixed benchmark profile. Myco won F1 at D = 10 and lagged the DE family on F2–F7; the paper attributes effectiveness to landscape structure and information transfer across local search regions, but does not yet identify which landscape properties predict success.
• Full disclosure and tuning of the ablation. The paper states that community structure and cord plasticity have distinct roles, but the underlying numerical ablation results are not available in the provided content; systematic sensitivity studies over K_max, λ_intra, λ_inter, η, δ, and τ_align are a natural next step.
• Filling out the dimensional comparison. Complete D = 20 results and how Myco's 30-tip configuration scales against the 50-individual comparators under the 50,000D budget remain to be examined.
• Biological fidelity versus algorithmic utility. The authors explicitly state that Myco does not model chemotropic growth or substrate-level physiology, leaving open how much further mycelial structure — beyond conductance reinforcement and mesoscale domains — could be exploited.

Target Audience

Researchers and graduate students in computational intelligence, metaheuristics, and nature-inspired computing who are interested in how interaction topology affects search; practitioners in engineering design and optimisation who need a budget-constrained continuous optimiser and want to understand its benchmark behaviour; and readers working on graph-based or network-oriented optimisation who want to see graph dynamics applied to a continuous problem rather than to a network problem.

Authors’ abstract

Continuous optimisation methods need to balance sharing information and maintaining alternative search directions. In this paper, we introduce Mycelial Search (Myco), a graph-structured metaheuristic designed around active tips, community-weighted flow, adaptive cord plasticity, and anchor-based injection. Candidate solutions form an evolving spatial graph in which a Louvain partition distinguishes within-community from cross-community information exchange. Adaptive cord plasticity subsequently modifies active tip-to-tip edges according to their alignment with the local flow. An anchor-based injection mechanism supplements the graph-driven tip dynamics. We evaluated Myco on the CEC 2022 single-objective bound-constrained benchmark suite at dimensions $D=10$ and $D=20$, using 30 independent runs per algorithm-function pair. The comparison includes eleven established optimisers from several search families. Myco reaches competitive results on selected functions across both dimensions. The ablation analysis further shows that community structure regulates the range of graph-based information exchange, whereas cord plasticity controls the persistence of local directional influence. These findings indicate that graph-structured local interaction can support continuous optimisation, while its effectiveness depends on landscape structure and information transfer across local search regions.