This paper presents a novel approach to the two-dimensional orthogonal rectangular strip packing problem (2D-SPP), focusing on arranging rectangular items on a strip of fixed width to minimize the total height. We extend the recently proposed novel neural-based constructive heuristic by integrating a local search strategy that explores the space of heuristics, encoded as neural network weights, rather than the solution space. Utilizing the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) for both training and problem-specific refinement, our method dynamically learns and optimizes placement decisions. We introduce benchmarking datasets inspired by real-world fast-moving consumer goods (FMCG) logistics, demonstrating the approach’s efficacy. Experimental results show that our refined constructive heuristic (RCH) significantly outperforms traditional heuristics like MaxRects and Skyline, and achieves up to tenfold reductions in wasted space compared to constraint programming for larger, heterogeneous instances. The proposed method offers scalability and adaptability, making it particularly suited for practical applications in logistics and manufacturing.

The Vehicle Routing Problem (VRP) has been extensively studied in the literature, leading to various extensions that incorporate real-world constraints. However, many practical considerations remain underrepresented, particularly in the FMCG (Fast-Moving Consumer Goods) industry. In this study, we analyze a real-world VRP formulation based on logistics operations from an FMCG distribution network and identify critical gaps between theory and practice. Based on our experience, we identify typical requirements in realworld applications that are relatively underrepresented in the existing literature. To demonstrate that the problem remains solvable, we propose a genetic algorithm (GA)-based approach, implemented within an open-source, object-oriented model, specifically designed to optimize route planning while incorporating real-world operational constraints. The algorithm was tested on a dataset covering 60 days of real-world logistics data, demonstrating significant improvements over historical benchmarks in terms of cost reduction, route efficiency, vehicle utilization, and operational feasibility. Our findings highlight the importance of bridging the gap between theoretical VRP formulations and industry-specific challenges, opening new avenues for further research and optimization in complex logistics networks.

This paper introduces a novel neural-based method for solving the Traveling Salesman Problem (TSP) by learning adaptive, single-pass construction heuristic within the Insertion heuristic framework. We design a simple feed-forward neural network that dynamically evaluates candidate cities based on multi-dimensional distance properties, allowing flexible handling of varying problem sizes. The network is trained using a black-box optimization approach with Covariance Matrix Adaptation Evolution Strategy (CMA-ES), bypassing the need for gradient information. To further improve solution quality, we propose a lightweight local refinement technique that perturbs the trained network’s weights, effectively exploring nearby heuristic landscapes without restarting the search process. Computational experiments on synthetic datasets with 50, 100, and 200 cities show that our approach consistently outperforms classical Insertion heuristics and generates solutions within 0–2.4% of optimality. The proposed method combines high computational efficiency with adaptability, demonstrating strong potential for extension to more complex TSP variants and practical logistics optimization tasks.

This paper introduces the concept of Conversational Decision Support Systems (C-DSS)-a novel, agent-based framework that leverages Large Language Models (LLMs) to enhance the Operations Research (OR) methodology.We focus on the modeling and coding stages of decision support systems, where language-based interaction is crucial.The paper evaluates the effectiveness of LLMs in generating mathematical models and AMPL code for a curated set of 20 LP/MILP artifacts.Four architectural setups are analyzed: a monolithic LLM agent (M/C), its enhancement with a code verifier (M/C+V), agentbased decomposition with RAG-enhanced coding (M+C R +V), and full specialization with RAG-enhanced modeling and coding (M R +C R +V).Experimental results on two benchmark problems reveal that the targeted retrieval-augmented generation technique (RAG) significantly improves performance for complex modeling patterns such as piecewise functions, indicator constraints, and nested logic.We also propose a broader vision of C-DSS as a multi-agent ecosystem-including agents for visualization, explanation, verification, and orchestration-suggesting a path toward more explainable, adaptable, and intelligent decision support systems.