Technical Summary: ZIVARI-TLBO

1. Problem Statement

Teaching–Learning-Based Optimization (TLBO) is a population-based optimizer that relies on teacher and learner phases. While grouped or multi-population extensions of TLBO can preserve distinct search trajectories, their effectiveness hinges on how information is exchanged between groups. Weak exchange isolates groups, while aggressive transfer can degrade diversity. A critical challenge in experimental fairness is that many migration operators generate new candidate solutions, requiring additional objective-function evaluations (calls). These calls consume the finite evaluation budget, making it difficult to isolate the value of the information-sharing mechanism itself from the cost of generating new candidates.

The paper addresses the need for an inter-group information-sharing mechanism that improves search dynamics without consuming the objective-function evaluation budget.

2. Methodology: ZIVARI-TLBO

ZIVARI-TLBO (Zero-Cost Inter-Group Evaluated-Elite Relay) is a specific implementation of a grouped TLBO framework (frozen as gts-v4-cm-fixed). It augments an existing grouped TLBO with a landscape-aware population-state controller by adding a fixed ring evaluated-elite relay.

Core Mechanism

• Topology: Groups are arranged in a fixed ring ($g \to (g+1) \mod G$).
• Relay Event: At scheduled intervals (every group epoch), the best already-evaluated learner (elite) from a source group is compared against the worst eligible learner in the target group.
• Selection Criteria:
  • The target excludes the current global-best position to prevent overwriting the best solution found so far.
  • If the source elite's stored objective value is better than the target's worst eligible learner, the source solution and its stored objective value replace the target.
  • If the source is not better, no change occurs.
• Zero-Cost Property: Crucially, the relay transfers an exact copy of an already-evaluated solution. It does not generate a new candidate, nor does it apply perturbation, noise, or interpolation. Consequently, the relay operation requires zero additional objective-function calls. The comparison and replacement rely entirely on previously computed fitness values.

Formal Definition

The method partitions the population $P(t)$ into $G$ disjoint groups. Let $e^{(t)}_g$ be the elite of group $g$ and $w^{(t)}_b$ be the worst eligible member of group $b$. The relay operation is defined as:
$$\text{if } f(e^{(t)}_a) < f(w^{(t)}_b): \quad w^{(t)}_b \leftarrow e^{(t)}_a$$
Since $f(e^{(t)}_a)$ is a stored value, $\Delta NFE_{relay} = 0$.

3. Key Contributions

The paper explicitly lists six contributions, emphasizing rigorous experimental design and statistical caution:

1. Fixed Ring Relay: A mechanism transferring source-group elites to the worst eligible learners of the next group using stored objective values.
2. Budget Audit: An implementation-level proof demonstrating that the exact relay operation incurs no additional objective-function calls.
3. Direct Ablation: A controlled comparison against the same grouped landscape-aware controller without the relay, using matched seeds and equal budgets.
4. Multidimensional Comparison: Evaluation against TLBO, MCTLBO, DE, PSO, GWO, WOA, and HHO across dimensions 10, 30, 50, and 100.
5. Statistical Analysis: Use of paired Wilcoxon signed-rank tests, Holm adjustment, rank-biserial effect sizes, and Friedman ranks.
6. Reproducibility: A package containing raw evidence, standardized data, scripts, and explicit claim limitations.

4. Experimental Results

Classical Benchmark Results

• Direct Ablation: Against the same controller without relay, ZIVARI-TLBO achieved 728 wins, 11 ties, and 221 losses across all dimensions. The rank-biserial effect size was 0.624, and the Holm-adjusted $p$-value was $1.411 \times 10^{-61}$. This confirms the relay improves performance within the tested protocol.
• Multidimensional Comparison: In a comparison of eight algorithms:
  • WOA achieved the best average rank (2.914).
  • ZIVARI-TLBO ranked second (3.382).
  • HHO ranked third (3.465) and was not significantly different from ZIVARI-TLBO after Holm adjustment.
  • ZIVARI-TLBO significantly outperformed TLBO, MCTLBO, DE, PSO, and GWO.
  • ZIVARI-TLBO lost significantly to WOA.

Engineering Design Results

• Evaluated on five constrained engineering problems (welded beam, pressure vessel, etc.) using a static quadratic penalty.
• Results were mixed and sensitive to constraint handling.
• ZIVARI-TLBO achieved the second-best overall penalized rank (behind DE).
• Feasibility rates varied significantly by problem (e.g., 30/30 feasible for tension-spring, 0/30 for three-bar truss).
• Under a Deb-style feasibility-first analysis, ZIVARI-TLBO ranked second (5.200) behind DE (4.933).

5. Significance and Claims

The paper adopts a modest and scoped stance regarding its significance:

• Validated Contribution: The evidence supports the claim that a scoped, zero-cost relay contributes positively to the performance of the specific grouped TLBO framework tested. It demonstrates that information sharing can occur without consuming the evaluation budget.
• No Universal Superiority: The authors explicitly reject claims of universal state-of-the-art performance, global convergence, or engineering dominance.
  • ZIVARI-TLBO did not rank first in the broad comparison (WOA did).
  • It is not significantly better than HHO.
  • Engineering results are mixed and dependent on the static-penalty formulation.
• Limitations: The study does not include verified implementations of modern state-of-the-art algorithms like CMA-ES, SHADE, L-SHADE, or jSO. It does not claim to solve the CEC benchmark suite.
• Conclusion: The primary value is the demonstration of a budget-consistent information-sharing mechanism and the rigorous isolation of its effect via ablation. The paper concludes that the appropriate next step is independent review and validation against modern baselines and verified CEC implementations, rather than further versioning of the optimizer.