International Journal of Advances in Applied Computational Intelligence

Reliable crop-yield prediction is essential for strategic food planning, input allocation, and precision-agriculture analytics, yet a single predictive model must accommodate substantial heterogeneity across crop species, regions, seasons, and management conditions. This paper proposes a hybrid computational-intelligence framework that couples extreme gradient boosting with an ensemble of residual neural networks in order to model both dominant nonlinear structure and systematic residual error. A leakage-controlled benchmark was constructed from a recent public agricultural data release, yielding 2,698 observations spanning eight major crops, 30 Indian states and union territories, six agricultural seasons, and ten complete years from 2010 to 2019. The predictor space combines climatic and agronomic descriptors, including annual rainfall, cultivated area, fertilizer and pesticide intensities, crop identity, state, season, temporal trend, and rainfall–management interaction terms; the production variable was removed to prevent algebraic target leakage. The empirical study compares the proposed framework with nine competitive baselines under a strict chronological design in which 2010–2017 are used for training, 2018 for model selection, and 2019 for independent testing. On the holdout year, the proposed hybrid attains R2 = 0.974, RMSE = 3.146 t ha−1, and MAE = 1.080 t ha−1, outperforming standalone XGBoost by 20.3% in RMSE and a direct deep neural network by 32.5%. Bootstrap analysis places the hybrid RMSE within 2.042–4.112 t ha−1 at the 95% confidence level. Grouped permutation analysis shows that crop identity, geographic location, and cultivated area contribute most strongly to predictive performance. Overall, the results support residual error correction as an effective strategy for strengthening tabular yield prediction while also underscoring the importance of leakage control, temporal validation, and competitive baseline design.

Synthetic industrial telemetry can alleviate scarcity and confidentiality constraints, but its value depends on more than distributional similarity. Generated sequences must support downstream engineering analysis, limit disclosure risk, and satisfy operational relations. This paper presents a reproducible three-axis benchmark and a differential-evolution calibration layer for conditional denoising diffusion. Six generators–bootstrap resampling, a shrinkage-Gaussian model, Fourier surrogates, a conditional variational autoencoder, a conditional diffusion model, and its calibrated counterpart–are evaluated on chronologically partitioned manufacturing telemetry. The protocol covers 18,000 generated 80-minute windows and combines train-synthetic/test-real alarm classification, marginal and temporal fidelity measures, record-proximity and membership-inference tests, and five engineering-rule audits. Differential evolution selects four post-generation controls using validation data only. Relative to the uncalibrated diffusion model, calibration reduces the aggregate physical-violation rate by 94.7% and the Wasserstein error by 29.2%, while increasing mean downstream ROC–AUC by 0.009. The improvement is accompanied by a 0.073 increase in membership-inference AUC and a small deterioration in autocorrelation error. Bootstrap resampling provides the strongest mean predictive utility but exactly reproduces 77.9% of its outputs; the conditional variational autoencoder attains the lowest Wasserstein error (0.046) with a 0.020% physical-violation rate. No generator dominates utility, privacy, and plausibility simultaneously. Synthetic industrial data should therefore be selected through deploymentspecific acceptance regions rather than a single realism score.

Large language models now routinely produce metaheuristic optimisation code in response to natural-language prompts, yet the structural integrity of such generated implementations has received little systematic scrutiny. Syntactically valid code that misrepresents algorithmic logic or parameter semantics constitutes a form of hallucination specific to optimisation software — one that is difficult to detect without execution-level testing and that can produce catastrophic performance degradation without any visible error signal. This work introduces a five-category taxonomy of algorithmic hallucination covering parameter value hallucination, update-formula omission, boundaryhandling failure, reproducibility failure, and termination-logic error. A controlled audit is conducted by injecting each hallucination type into verified reference implementations of three widely used metaheuristics — particle swarm optimisation, differential evolution, and a genetic algorithm — and evaluating their behaviour across five standard continuous benchmark functions. Across twenty independent trials per condition, parameter hallucination and formula-omission errors produce best-fitness degradation exceeding 108-fold relative to the reference, while boundary-handling failures generate more than 4 000 out-of-bounds constraint violations per trial. Reproducibility failure inflates inter-trial variance by three to four orders of magnitude, and premature-termination errors induce a 100% trial failure rate. The results demonstrate that current LLM-generated metaheuristic code requires structured execution-level validation before deployment, and a practical audit protocol is proposed to support that process. Experimental artefacts, including all source code, generated data, and benchmark results, are released to support reproducible follow-on research.

Metaheuristic optimization is extensively used in electrical and industrial engineering, where nonlinear, nonconvex, mixed-variable, and simulation-based models arise in power-system operation, energy management, production planning, scheduling, and manufacturing-system design. The rapid proliferation of newly named optimizers, however, makes it difficult to determine whether an apparent contribution introduces a distinct search mechanism or repackages an established algorithmic lineage. This article presents Metaheuristic Lineage Mining (MLM), a multimodal screening framework that integrates executable equation structure, normalized source-code structure, and empirical search dynamics. Identifier-invariant abstract-syntax motifs represent update equations, normalized token n-grams characterize implementation structure, and trajectory fingerprints summarize convergence, diversity, exploration, and improvement behavior. The evaluation includes 57 implementations grouped into 16 documented variant families and 1,710 controlled runs over six continuous benchmark functions and five random seeds. Fusion weights are selected under lineage-held-out validation, preventing algorithms from the evaluated family from influencing model selection. Code similarity is the strongest individual channel, with a mean ROC–AUC of 0.955 and top-1 lineage retrieval of 0.930; equation similarity reaches a ROC–AUC of 0.929. Search-dynamics similarity is weaker alone but supplies behaviorally independent corroboration. Constrained multimodal fusion yields the highest average precision of 0.769, a mean reciprocal rank of 0.935, and clustering normalized mutual information of 0.885. Interpretable cross-family associations, including WhaleFOA–WOA and JADE–SHADE, demonstrate the ability to expose hybrid or ancestral relations that categorical labels omit. The framework is not a detector of misconduct; it is an auditable engineering due-diligence tool for evaluating optimizer originality before costly application studies in power, energy, manufacturing, logistics, and production systems.