Technical Summary: Accumulated-History Replicator Selection Optimization (Mycelial Network Optimization)

1. Problem Statement

Population-based optimization algorithms (e.g., Particle Swarm Optimization, Differential Evolution) typically employ a "memoryless" selection mechanism. In these systems, a candidate solution's survival into the next generation is determined solely by its instantaneous fitness at time $t$. This design discards all historical performance data, treating an agent that has consistently occupied the top decile of the population for fifty iterations identically to a newly generated agent with a marginally better current fitness.

The paper argues that this temporal blindness forfeits a critical signal: the persistence of quality over time. While instantaneous fitness is sufficient for smooth, separable landscapes, it fails to capture the geometric information inherent in structured, non-separable landscapes where past performance indicates long-term viability. The authors posit that biological selection systems exploit this persistence, and that current metaheuristics lack a formally grounded mechanism to do so.

2. Methodology: Mycelial Network Optimization (MNO)

The paper introduces the Accumulated-History Replicator Selection (AHRS) framework, implemented as the Mycelial Network Optimization (MNO) algorithm. Inspired by fungal mycelia that accumulate cytoplasmic resources over time and share them via anastomosis, MNO replaces instantaneous fitness with an accumulated resource variable as the primary survival criterion.

Core Components

1. Cytoplasmic Streaming (Accumulated-History Selection):
   Each agent $i$ maintains a resource variable $R_i(t)$, updated as an exponentially weighted moving average of its normalized quality history $Q_i(t)$:
   $$R_i(t+1) = (1-\gamma)R_i(t) + \gamma Q_i(t)$$
   where $\gamma$ is the streaming rate. Agents are removed if their accumulated resource falls below a budget-dependent threshold. This ensures that survival depends on the complete trajectory of quality, not just the current state.

2. Fitness-Proximity Anastomosis:
   Every third iteration, agents with similar fitness values (within a niche threshold $\delta_a$) undergo resource transfer. The "winner" receives a fraction of the "loser's" resources, followed by a greedy chemotactic repositioning step toward the winner. This facilitates resource flow between agents occupying similar niches.

3. Natural-Gradient Step:
   Movement is guided by a resource-weighted natural-gradient step. The search ellipse is elongated along high-variance dimensions (productive valleys or ridges) based on the distribution of resource weights $p_i = R_i / \sum R_j$.

4. Sporulation:
   If the population drops below half its size, new agents are introduced (60% random, 40% opposition-based) to maintain diversity.

3. Theoretical Grounding

The paper provides a rigorous mathematical foundation for MNO, distinguishing it from purely metaphorical bio-inspired algorithms:

• History Dependence: Theorem 1 proves that the resource variable $R_i(t)$ is a sufficient statistic for an agent's complete past quality trajectory. Consequently, the selection criterion is non-memoryless; two agents with identical instantaneous fitness but different histories have different survival probabilities.
• Replicator Equation Connection: Theorem 2 demonstrates that the continuous-time limit of the streaming update corresponds exactly to the Replicator Equation from evolutionary game theory:
  $$\frac{dp_i}{dt} = \gamma p_i (Q_i - \bar{Q})$$
  The discrete update is shown to be the Euler discretization of this equation with a local truncation error bounded by $O(\gamma^2)$.
• Convergence Guarantee: By applying Fisher's Fundamental Theorem to the Replicator Equation, the paper proves that the resource-weighted mean quality $\bar{Q}(t)$ is non-decreasing in expectation throughout the run.

4. Experimental Results

MNO was evaluated against seven baselines (PSO, DE, SHADE, L-SHADE, JADE, ABC, CMA-ES) across:

• 20 classical benchmark functions.
• 15 CEC 2017 functions (D=30, D=50).
• 6 constrained engineering design problems.

Key Findings

• Structured vs. Separable Landscapes: MNO significantly outperforms PSO and DE on non-separable, structured functions (e.g., Rosenbrock, Schwefel 1.2, Trigonometric). For instance, on Rosenbrock (30D), MNO achieved a result ~1,400 times better than PSO.
• Performance on Smooth Functions: On smooth, separable functions (e.g., Sphere, Ackley), MNO underperforms compared to parameter-adaptive DE variants (SHADE, JADE, L-SHADE). The authors attribute this to the "No Free Lunch" theorem, noting that mechanisms exploiting temporal structure do not offer advantages on landscapes lacking such structure.
• Ablation Study: Experiments confirmed that both the streaming operator and the anastomosis operator contribute independently to performance. The combined algorithm improved 15/20 classical functions, exceeding the performance of either operator alone.
• Engineering Applications: MNO achieved results comparable to state-of-the-art on spring design (0.4% gap) and economic load dispatch (0.01% gap). A larger gap (23.6%) on the welded-beam problem was attributed to limitations of the exterior penalty method used for constraint handling, not the selection mechanism itself.

5. Significance and Claims

The paper makes a modest but specific claim regarding its novelty and significance:

• Novelty: The contribution is not the use of exponential smoothing (which exists in fitness inheritance) nor the Replicator Equation itself. The novelty lies in the specific combination of using an individually accumulated exponential quality history as the direct survival criterion, with its dynamics formally connected to evolutionary game theory via the Replicator Equation.
• Theoretical Contribution: MNO is the first metaheuristic to explicitly formulate its individual-level selection mechanism as a discretization of the Replicator Equation with formally verified error bounds ($O(\gamma^2)$).
• Practical Implication: The results identify "accumulated history" as a previously unexploited selection signal that is particularly effective on structured, non-separable search landscapes. The paper suggests that other population-based algorithms could benefit from incorporating temporal memory into survival decisions, particularly where landscape geometry is complex.

The authors explicitly acknowledge that MNO is not a universal solver superior to all existing methods; rather, it offers a theoretically grounded alternative for specific classes of optimization problems where temporal persistence of quality is a valid signal.