Energy Valley Optimizer (EVO)

• Energy Valley Optimizer (EVO) is a metaheuristic algorithm that mimics nuclear decay processes to efficiently navigate high-dimensional feature selection challenges.
• It employs multiple update regimes—similar to alpha, beta, and gamma decay—to dynamically balance exploration and exploitation in search spaces.
• Empirical results in intrusion detection demonstrate that EVO reduces feature dimensionality while achieving high classification accuracy and robust performance.

The Energy Valley Optimizer (EVO) is a population-based metaheuristic optimization algorithm, fundamentally inspired by concepts from nuclear and particle physics, particularly the tendency of unstable atomic particles to decay toward lower-energy, stable configurations. EVO has been explicitly developed to address high-dimensional, combinatorial search spaces, with notable applications in feature selection for supervised classification systems, such as hybrid intrusion detection systems (HyIDS) for cloud security (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026).

1. Theoretical Foundations and Physical Analogy

EVO is rooted in a physics-based metaphor where each solution candidate is regarded as a "particle" situated in an abstract "energy valley." This conceptualization aligns each algorithmic operator with a physically meaningful principle—such as the energy minimization trajectory observed in the decay behavior of neutrons and protons in atomic nuclei. The dynamics of the algorithm are directly mapped to mechanisms analogous to alpha, beta, and gamma decay, embodying both exploration (global search) and exploitation (local search), with three or four concurrent position-update regimes reflecting distinct physical transition processes (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026).

Key elements in the analogy include:

• Energy Landscape: The fitness surface is conceptualized as an energy landscape, with lower fitness corresponding to lower energy (i.e., more optimal configurations).
• Particle Interactions: Both local (neighbor-based) and global (population center, best-known solution) interactions guide the search process, mirroring gravitational attractions and local clustering observed in physical systems.
• Stability Levels and Energy Barriers: Transitions between exploration and exploitation regimes are determined by energy slopes and thresholds, analogous to overcoming physical barriers to transition between states.

2. Mathematical Formulation and Update Equations

Formally, EVO operates with a population of $n$ particles $X_i$ in a $D$-dimensional space ($X_i\in \{0,1\}^D$ for binary feature selection or $X_i\in\mathbb{R}^D$ for continuous domains). The objective is to minimize a fitness function $f(X)$, commonly defined as a function of classifier accuracy, false positive rate (FPR), and false negative rate (FNR) (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026).

EVO uses multiple update rules corresponding to decay analogies:

• Alpha-decay (Drift to Best)

$$X_i^{\text{new}} = X_i + \alpha (X_{\mathrm{best}} - X_i)$$

• Gamma-decay (Local Perturbation)

$$X_i^{\text{new}} = X_i + \gamma (X_{\mathrm{NG}} - X_i)$$

• Beta-decay I (Best & Population Center, scaled by Slope)

$$X_i^{\text{new}} = X_i + \frac{T_1(X_{\mathrm{best}} - X_i) - T_2 (X_{\mathrm{CP}} - X_i)}{S_{L,i}}$$

• Beta-decay II (Best–Neighbor Exchange)

$$X_i^{\text{new}} = X_i + T_3 (X_{\mathrm{best}}-X_i) - T_4(X_{\mathrm{NG}}-X_i)$$

Here,

• $X_{\mathrm{best}}$ is the global best solution,
• $X_{\mathrm{CP}} = \frac{1}{n}\sum_{j=1}^n X_j$ is the population center,
• $X_{\mathrm{NG}}$ is a cluster or nearest neighbor,
• $S_{L,i}$ encodes the local energy slope,
• $T_k \sim U(0,1)$ are randomized coefficients, and parameters $\alpha, \gamma$ are set accordingly.

Regime selection (the choice among these updates) is driven by the relative "energy" (fitness) of each particle and the dynamically computed energy barrier (typically $EB = \alpha \max_j f(X_j)$ with $\alpha \in (0,1)$).

3. Algorithmic Workflow and Pseudocode

EVO proceeds as follows (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026):

1. Initialization: Population of binary (feature-presence) or continuous vectors randomly sampled within specified bounds.
2. Fitness Evaluation: For each particle, classifier-based fitness is computed—either as negative cross-validation accuracy or a weighted sum of accuracy, FPR, and FNR:

$$\mathrm{Cost}(X) = W_1(1 - \mathrm{Accuracy}(X)) + W_2\,\mathrm{FPR}(X) + W_3\,\mathrm{FNR}(X)$$

1. Population Dynamics (Iteration):
   • Compute Population Center.
   • Cluster/Neighbor Identification: Compute pairwise (often Hamming) distances to identify nearest neighbors and cluster centroids.
   • Update Rule Selection: If $f(X_i) < EB$, use a regime emphasizing exploitation; otherwise, emphasize exploration.
   • Apply Selected Decay Rule with random or slope-adaptive coefficients.
   • Clipping/Repair: Ensure all updated particles respect variable domain constraints, e.g., binary coding enforced by thresholding sigmoid outputs.
   • Evaluate and Update Best: Fitness recalculation and best-solution update as required.

A simplified high-level pseudocode (binary case) is as follows:

Initialize {X_i}₁ⁿ randomly in {0,1}^D Evaluate f(X_i); set X_BS ← argmin_i f(X_i) while not termination: for i in 1…n: find two nearest neighbors of X_i by distance X_NG ← round((nbr1 + nbr2)/2) X_CP ← round((1/n)∑_j X_j) EB ← α·max_j f(X_j) if f(X_i) < EB: X_i ← round[X_i + r₁·(X_BS−X_i) + r₂·(X_CP−X_i)] else: X_i ← round[X_NG + r₃·(X_BS−X_NG)] clip X_i to {0,1} evaluate f(X_i) if f(X_i) < f(X_BS): X_BS ← X_i return X_BS

where $r_k \sim U(0,1)$.

4. Parameters, Complexity, and Convergence Properties

Critical algorithm parameters include:

• Population size ($n$): 30–50 typical, larger values increase diversity but incur $O(n^2)$ time per iteration due to distance calculations.
• Maximum evaluations (MaxFes/MaxIter): 100–500 used in practical deployments.
• Energy barrier coefficient $\alpha$: 0.6–0.9 for controlling exploration vs. exploitation trade-off.
• Decay coefficients $r_k, T_k$: random variables in [0,1], recalculated each step.

The dominant cost per iteration is pairwise distance computation ($O(n^2)$), plus $O(n\cdot \text{cost(fitness)})$ for classifier training and validation. Overall, computational complexity per run is approximately $O(\text{MaxFes} \cdot (n^2 + \text{CostEval}))$ (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026).

Convergence in practice is observed within a few dozen iterations, driven by the explicit exploration/exploitation switch and decay-mode adaptivity, with empirical evidence of robust avoidance of local minima.

5. Performance Comparison and Empirical Results

EVO's application as a feature selection wrapper in HyIDS systems has been evaluated on multiple intrusion detection datasets, including CIC_DDoS2019, CSE_CIC_DDoS2018, and NSL-KDD (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026). The optimizer was compared against the Grey Wolf Optimizer (GWO) across classifiers such as Decision Tree, SVM, Random Forest, and KNN.

Key results:

• Dimensionality Reduction:
  • CIC_DDoS2019: 88 → 38 features (Alhusseini et al., 3 Jan 2026), 87 → 81 features (Al-Husseini, 24 Jun 2025).
  • CSE_CIC-IDS2018: 80 → 43 (Alhusseini et al., 3 Jan 2026), 80 → 78 (Al-Husseini, 24 Jun 2025).
  • NSL-KDD: 42 → 41 (Al-Husseini, 24 Jun 2025).
• Detection Accuracy:
  • D_TreeEVO (CIC_DDoS2019): 99.13% accuracy, 98.94% F1 score (Alhusseini et al., 3 Jan 2026).
  • D_TreeEVO (CSE_CIC-IDS2018): 99.78% accuracy, 99.70% F1 score (Alhusseini et al., 3 Jan 2026).
  • Outperforms no-feature-selection and GWO-based feature selection in nearly all settings (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026).
• Computational Cost: Typical EVO+Decision Tree runtime <1 s per dataset; convergence in ≈100 evaluations (Al-Husseini, 24 Jun 2025).

Dataset	Features (before→after)	D-TreeEVO Accuracy	F1/Detection Rate
CIC_DDoS2019	88 → 38	99.13%	98.94%
CSE_CIC-IDS2018	80 → 43	99.78%	99.70%
NSL-KDD	42 → 41	99.50%	99.47%

EVO-based selection consistently reduces feature dimensionality while maintaining or improving classifier performance relative to both baseline (all features) and GWO (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026).

6. Integration into Hybrid Intrusion Detection and Practical Considerations

In the HyIDS pipeline, EVO is deployed as a wrapper feature-selection technique. Each particle encodes a subset of features; performance is scored using a generic weighted cost function reflecting multiple classification objectives (accuracy, FPR, FNR). Post-selection, the classifier is trained and evaluated on the reduced feature set.

The practical workflow is:

1. Data preprocessing (downsampling, scaling, encoding).
2. Application of EVO for feature selection.
3. Training classifiers (SVM/RF/D-Tree/KNN) on the reduced set.
4. Evaluation on held-out test subsets.

This hybrid structure allows flexible and modular optimization across algorithmic components and is effective even under severe class imbalances, as demonstrated in both the CIC_DDoS2019 and CSE_CIC-IDS2018 cases (Alhusseini et al., 3 Jan 2026).

7. Limitations, Insights, and Prospective Directions

• Insights:
  • EVO achieves explicit, interpretable control over exploration/exploitation via the energy barrier parameter.
  • Strong conceptual–model alignment with physical principles yields stable performance across diverse, high-dimensional datasets.
  • Empirically, the algorithm shows best relative improvement with shallow base classifiers; benefits on Random Forest are less pronounced (Al-Husseini, 24 Jun 2025).
• Limitations:
  • $O(n^2)$ complexity for neighbor discovery is a significant limiting factor for large populations.
  • Feature mixing may be slow in high-dimensional, binary-encoded search spaces.
  • Performance is sensitive to parameter tuning, particularly the population size and energy barrier coefficient (Al-Husseini, 24 Jun 2025).
• Future Directions:
  • Adaptive or self-tuning of parameters for dynamic energy barrier and population.
  • Metaheuristic hybridization (e.g., EVO–GWO or EVO–PSO compositions).
  • GPU-accelerated implementations for large-scale use cases.
  • Multi-objective extensions balancing accuracy and feature subset parsimony (Al-Husseini, 24 Jun 2025, Alhusseini et al., 3 Jan 2026).

EVO offers a novel paradigm for combinatorial optimization in machine learning pipelines, especially where interpretability of the search process and high-fidelity analogy to physical systems are desired. Its adoption in IDS and feature selection showcases both flexible applicability and promising empirical performance over established alternatives.

---

References

• (Al-Husseini, 24 Jun 2025)
• (Alhusseini et al., 3 Jan 2026)