Multi-Agent Systems (MAS) can benefit from Large Language Models (LLMs), but hallucinations pose risks to decision-making. This paper introduces an approach for creating MAS based on LLMs and proposes a generalized architecture for such systems. We ensure that reasoning is conducted through predicate logic to minimize errors, and LLMs are exclusively utilized to translate natural language into Prolog, supported by our algorithm for generated code correction. In addition, a safety reasoning agent is introduced to validate facts and prevent rule violations. We evaluate this approach in a sample MAS, demonstrating accurate logical transformations. Results confirm that LLMs add value to MAS when combined with structured knowledge representation, offering opportunities for further exploration focused on scalability enhancement, optimizing efficiency and dynamic refining of ontologies.
This paper deals with the loop machine layout problem. From the classical perspective, it is considered as a quadratic assignment problem (QAP) where we seek to order machines on prefixed positions forming a loop layout in order to minimize the total cost of transporting parts within each manufacturing cell. In the present paper, another formulation of the same problem is exhibited that take into account proximity constraints between machines. Hence, during this study, we aim to construct a path between already placed machines in order to form a loop layout with a single shortcut. This processing makes it look similar to the Travelling Salesman Problem (often called TSP) with lightly difference version. Herein, to solve this problem, we propose an improved genetic algorithm using a new crossover operator named Precedence Crossover (PCX). The choice of an effective crossover operator has a major impact on the quality of the near-optimal solutions. Many works have proposed different crossover operators but none of them is assigned to solve the new machine layout problem formulation. Comparing to many existing operators, experimental results reveal that proposed one provides better results on several levels.
It is important to analyze online data whose characteristics change over time, and many studies have been conducted. The change in characteristics over time is called concept drift. In general, ensemble learning methods, such as ARF and SRP, perform well as analysis methods. Our previous paper reported an adverse effect of a characteristic that has yet to be considered in the analysis of concept drifts and proposed a solution. The characteristic discussed in our previous paper is the ratio of explanatory variables whose characteristics change. The previous paper shows that when this ratio is large, it harms the ensemble learning methods. In addition, it proposed SARF as a solution and reported that SARF could reduce the misclassification rate of ARF on 20 data out of 26 public data. The key idea of SARF is the selection of ARF models with the best starting points. This paper further enhances SARF by introducing a voting mechanism. The voting mechanism reduces SARF’s misclassification rate by using information from past data, which SARF ignores. This paper also shows that the idea of “selecting and voting learning models with best starting points” can enhance other ensemble learning methods. Voting mechanisms by the models with the best starting points can reduce the misclassification rate of both ARF and SRP on 29 out of 34 public data.
This paper introduces new methods for enhancing effort estimation by integrating project, effort, and industry dimensions. Two collaborative filtering principles are utilized: a neighbor-based approach, which estimates effort based on similar projects within specific industry sectors, and a latent factor-based model that optimizes an objective function through regularization and positive constraints. Experimental results demonstrate the effectiveness of both methods, with a statistical comparison indicating that the latent factor model consistently outperforms the neighbor-based approach by achieving lower RMSE values. These findings highlight the ability of the latent factor model to capture complex relationships within software project data, contributing to better industry outcomes.
With the rapid growth of recommendation systems and the increasing demand for efficient handling of high-dimensional data, vector database management systems (VDBMS) have become a critical component in modern machine learning and AI-driven applications. This survey provides a comprehensive review of VDBMS from the perspective of recommendation systems, analyzing their architectures, indexing techniques, query processing strategies, and scalability. We explore how vector databases optimize nearest neighbor search, similarity computation, and retrieval performance, which are essential for recommendation tasks. Additionally, we compare popular VDBMS solutions, highlighting their advantages and trade-offs in different recommendation scenarios. Finally, we discuss emerging trends and challenges in the field, including hybrid models, real-time updates, and integration with deep learning techniques. This study aims to guide researchers and practitioners in selecting and optimizing VDBMS for next-generation recommendation systems.
