Interactive multi-criterion decision-making (iMCDM) procedures generally incorporate the human decision maker’s (DM’s) preferences to iteratively compute preferred Pareto-optimal (PO) solutions. DM preferences often involve objective classification indicating their improvement, relaxation, satisfaction, and indifference at the current PO solution. In addition, the DM provides one or more preference information in the form of a reference point, objective weights, and reference direction. These preferences are used to reformulate the original multi-objective optimization (MOO) problem into a single-objective optimization problem using the achievement scalarization function (ASF). The iterative preferences are then used to solve the ASF problem to compute a preferred PO solution. In this paper, we propose a modified version of a well-known iMCDM procedure: GUESS. In order to compare the modified GUESS with the original GUESS, we first propose a Machine Learning (ML)–based decision-maker (Machine-DM), which uses two Artificial Neural Networks (ANNs) to replace human DM steps in GUESS. The two trained ANNs predict the objective class and bounding parameters of objectives from the decision variable vector, without explicit information on the location and value of the preferred point, thereby enabling a fair performance evaluation. Following the Machine-DM approach, three variants of GUESS: MachDM-GUESS (original), MachDM-GUESS-I and MachDM-GUESS-II (modified), are developed and compared using newly proposed performance metrics. The implementation of Machine-DM with the original and modified GUESS procedures reveals superiority of one of the modified GUESS procedures. Moreover, the study conducted in this paper opens up new avenues for benchmarking existing iMCDM procedures and proposing new and competitive ones, paving the way and encouraging EMO and MCDM researchers to conduct more such studies.
Large Neighborhood Search (LNS) and 2-opt local search are key algorithms for routing problems like TSP and CVRP. Standard hybrid frameworks usually separate these steps, using LNS for global exploration and 2-opt for later optimization. This order often leads to fault propagation, where suboptimal edges from repair are only fixed after costly global scans. To address this limitation, we investigate a Process-Aware Window Mechanism for 2-opt based LNS. Unlike other granular search methods that use fixed geometric metrics, our method uses the dynamic information from the repair operator. We set an optimization ’window’ around the insertion index of new nodes. This change makes 2-opt a proactive, built-in correction tool instead of just a post-processing step. We test six window selection strategies, from static to adaptive. Our experimental analysis shows that this frequent, low-complexity micro-optimization can improve convergence speed and result stability compared to traditional global search methods.
This paper presents a systematic comparison of spatial interpolation methods for reconstructing missing solar irradiance data on quasi-regular latitude–longitude grids. Using PVGIS-derived datasets, we evaluate interpolation performance under realistic missing-data scenarios, including (M1) random removal, (M2) contiguous gaps, and (M3) regular-grid thinning. A reproducible evaluation pipeline is used to apply k-nearest neighbors, inverse distance weighting, and analytical grid-based methods (linear, nearest) across multiple regions, orientations, and hyperparameter settings. Reconstruction accuracy is assessed using standard pointwise error metrics. Across all scenarios, K-Nearest Neighbors consistently outperforms alternative methods, demonstrating lower reconstruction errors and higher explanatory power even at increased levels of data removal. The results highlight method-dependent performance trends and trade-offs under different missing-data patterns. The results provide practical guidance for interpolation method selection in solar resource assessment and renewable energy applications.
The Electric Vehicle Charging Scheduling Problem (EVCSP) concerns the allocation of limited charging resources to heterogeneous vehicle requests under temporal and capacity constraints. Although advanced optimization models have been proposed for this problem, most assume an offline setting with complete knowledge of future arrivals, limiting their applicability in real charging stations. This paper introduces an online, preemptive EVCSP that preserves a high-fidelity problem formulation, including charger availability and grid capacity constraints, while enforcing a commitment policy under which admission decisions are irrevocable. We propose a hybrid solution approach that combines a genetic algorithm for online admission control and charger assignment with a mathematical optimization model for energy allocation. To meet real-time decision-making requirements, we embed a Gaussian-process surrogate within the evolutionary search to reduce the cost of repeatedly evaluating candidate schedules. Computational experiments across multiple instance sizes and booking modes show that surrogate assistance preserves objective values relative to exact evaluation while yielding statistically significant reductions in computational time, supporting practical online deployment of evolutionary optimization for real-time EV charging station operations.
Surrogate-assisted evolutionary algorithms and Bayesian optimization are widely used to reduce the cost of hyperparameter optimization (HPO) under limited evaluation budgets. However, under tight budgets, a surrogate can systematically overestimate the true loss in some regions, causing false discards: potentially improving configurations are rejected by conservative screening rules and never receive expensive evaluations. We propose ARC-CTRL, an audit-guided closed-loop trust controller that wraps a surrogate-assisted evolutionary HPO pipeline with an operational reliability layer. ARC-CTRL turns surrogate trust failures into an online observable binary event—miscoverage, defined as a selection-time lower-bound violation y < L(x)—and uses a Beta–Binomial posterior upper credible bound Un as an operational audit signal. A multiplicative feedback controller adapts a conservatism parameter γ to drive the audit signal Un toward a target risk level αrisk. ARC-CTRL is surrogate-agnostic: it does not require posterior variance; instead it builds local risk proxies from cross-fitted residuals and kNN local quantiles in mixed discrete–continuous HPO spaces. On three Hugging Face text classification tasks (Emotion, AG News, SST-2) under budgets B ∈ {40,60} with 11 paired seeds, ARC-CTRL consistently reduces miscoverage versus a strong fixed-γ control (LRQ-FIX) while satisfying a pre-specified non-inferiority criterion on best validation loss. Un is a Bayesian credible bound used operationally; we do not claim distribution-free guarantees under adaptive sampling.
