All issues

This paper investigates the effectiveness of Retrieval-Augmented Generation (RAG) combined with various Large Language Models (LLMs) for document retrieval and information access in corporate information systems. We survey typical use-cases of LLMs in enterprise environments, outline the RAG architecture, and discuss the major challenges that arise when integrating LLMs into a RAG pipeline. A system architecture is proposed that couples a text-vector encoder with an LLM. The encoder builds a vector database that indexes a library of corporate documents. For every user query, relevant contextual fragments are retrieved from this library via the FAISS engine and appended to the prompt given to the LLM. The LLM then generates an answer grounded in the supplied context. The overall structure and workflow of the proposed RAG solution are described in detail. To justify the choice of the generative component, we benchmark a set of widely used LLMs — ChatGPT, GigaChat, YandexGPT, Llama, Mistral, Qwen, and others — when employed as the answer-generation module. Using an expert-annotated test set of queries, we evaluate the accuracy, completeness, linguistic quality, and conciseness of the responses. Model-specific characteristics and average response latencies are analysed; the study highlights the significant influence of available GPU memory on the throughput of local LLM deployments. An overall ranking of the models is derived from an aggregated quality metric. The results confirm that the proposed RAG architecture provides efficient document retrieval and information delivery in corporate environments. Future research directions include richer context augmentation techniques and a transition toward agent-based LLM architectures. The paper concludes with practical recommendations on selecting an optimal RAG–LLM configuration to ensure fast and precise access to enterprise knowledge assets.

In this paper, we propose an adaptive algorithm that simulates the process of forming the initial behavioral skills on the example of the system ‘eye-arm’ animat. The situation is the formation of the initial behavioral skills occurs, for example, when a child masters the management of their hands by understanding the relationship between baseline unidentified spots on the retina of his eye and the position of the real object. Since the body control skills are not ‘hardcoded’ initially in the brain and the spinal cord at the level of instincts, the human child, like most young of other mammals, it is necessary to develop these skills in search behavior mode. Exploratory behavior begins with trial and error and then its contribution is gradually reduced as the development of the body and its environment. Since the correct behavior patterns at this stage of development of the organism does not exist for now, then the only way to select the right skills is a positive reinforcement to achieve the objective. A key feature of the proposed algorithm is to fix in the imprinting mode, only the final action that led to success, and that is very important, led to the familiar imprinted situation clearly leads to success. Over time, the continuous chain is lengthened right action — maximum use of previous positive experiences and negative ‘forgotten’ and not used.

Thus there is the gradual replacement of the random search purposeful actions that observed in the real young. Thus, the algorithm is able to establish a correspondence between the laws of the world and the ‘inner feelings’, the internal state of the animat. The proposed animat model was used 2 types of neural networks: 1) neural network NET1 to the input current which is fed to the position of the brush arms and the target point, and the output of motor commands, directing ‘brush’ manipulator animat to the target point; 2) neural network NET2 is received at the input of target coordinates and the current coordinates of the ‘brush’ and the output value is formed likelihood that the animat already ‘know’ this situation, and he ‘knows’ how to react to it. With this architecture at the animat has to rely on the ‘experience’ of neural networks to recognize situations where the response from NET2 network of close to 1, and on the other hand, run a random search, when the experience of functioning in this area of the visual field in animat not (response NET2 close to 0).

Automated urban traffic monitoring systems are widely used to solve various tasks in intelligent transport systems of different regions. They include video enforcement, video surveillance, traffic management system, etc. Effective traffic management and rapid response to traffic incidents require continuous monitoring and analysis of information from these complexes, as well as time series forecasting for further anomaly detection in traffic flow. To increase the forecasting quality, data fusion from different sources is needed. It will reduce the forecasting error, related to possible incorrect values and data gaps. We implemented the approach for short-term and middle-term forecasting of traffic flow (5, 10, 15 min) based on data fusion from video enforcement and video surveillance systems. We made forecasting using different recurrent neural network architectures: LSTM, GRU, and bidirectional LSTM with one and two layers. We investigated the forecasting quality of bidirectional LSTM with 64 and 128 neurons in hidden layers. The input window size (1, 4, 12, 24, 48) was investigated. The RMSE value was used as a forecasting error. We got minimum RMSE = 0.032405 for basic LSTM with 64 neurons in the hidden layer and window size = 24.

The paper considers the process of flow formation in an outflow of a binary gas mixture through a thin slit into vacuum. An approach to modeling the flows of rarefied gas mixtures in the transient regime is proposed based on the direct solution of the Boltzmann kinetic equation, in which the conservative projection method is used to calculate the collision integrals. Calculation formulas are provided; the calculation procedure is described in detail in relation to the flow of a binary gas mixture. The Lennard–Jones potential is used as an interaction potential of molecules. A software modeling environment has been developed that makes it possible to study the flows of gas mixtures in a transitional regime on systems of cluster architecture. Due to the use of code parallelization technologies, an acceleration of calculations by 50–100 times was obtained. Numerical simulation of a two-dimensional outflow of a binary argon-neon gas mixture from a vessel into vacuum through a thin slit is carried out for various values of the Knudsen number. The graphs of the dependence of gas mixture components output flow on time in the process of establishing the flow are obtained. Non-stationary regions of strong separation of gas mixture components, in which the molecular densities ratio reaches 10 or more, were discovered. The discovered effect can have applications in the problem of gas mixtures separation.

Deep learning’s power stems from complex architectures; however, these can lead to overfitting, where models memorize training data and fail to generalize to unseen examples. This paper proposes a novel probabilistic approach to mitigate this issue. We introduce two key elements: Truncated Log-Uniform Prior and Truncated Log-Normal Variational Approximation, and Automatic Relevance Determination (ARD) with Bayesian Deep Neural Networks (BDNNs). Within the probabilistic framework, we employ a specially designed truncated log-uniform prior for noise. This prior acts as a regularizer, guiding the learning process towards simpler solutions and reducing overfitting. Additionally, a truncated log-normal variational approximation is used for efficient handling of the complex probability distributions inherent in deep learning models. ARD automatically identifies and removes irrelevant features or weights within a model. By integrating ARD with BDNNs, where weights have a probability distribution, we achieve a variational bound similar to the popular variational dropout technique. Dropout randomly drops neurons during training, encouraging the model not to rely heavily on any single feature. Our approach with ARD achieves similar benefits without the randomness of dropout, potentially leading to more stable training.

To evaluate our approach, we have tested the model on two datasets: the Canadian Institute For Advanced Research (CIFAR-10) for image classification and a dataset of Macroscopic Images of Wood, which is compiled from multiple macroscopic images of wood datasets. Our method is applied to established architectures like Visual Geometry Group (VGG) and Residual Network (ResNet). The results demonstrate significant improvements. The model reduced overfitting while maintaining, or even improving, the accuracy of the network’s predictions on classification tasks. This validates the effectiveness of our approach in enhancing the performance and generalization capabilities of deep learning models.

For cognitive radio systems, it is important to use efficient algorithms that search for free channels that can be provided to secondary users. Therefore, this paper is devoted to improving the accuracy of prediction frequency resource occupancy of a cellular communication system using spatiotemporal radio environment maps. The formation of a radio environment map is implemented for the fourthgeneration cellular communication system Long-Term Evolution. Taking this into account, a model structure has been developed that includes data generation and allows training and testing of an artificial neural network to predict the occupancy of frequency resources presented as the contents of radio environment map cells. A method for assessing prediction accuracy is described. The simulation model of the cellular communication system is implemented in the MatLab. The developed frequency resource occupancy prediction model is implemented in the Python. The complete file structure of the model is presented. The experiments were performed using artificial neural networks based on the Long Short-Term Memory and Kolmogorov – Arnold neural network architectures, taking into account its modification. It was found that with an equal number of parameters, the Kolmogorov –Arnold neural network learns faster for a given task. The obtained research results indicate an increase in the accuracy of prediction the occupancy of the frequency resource of the cellular communication system when using the Kolmogorov – Arnold neural network.

The short-term prediction of road traffic condition is one of the main tasks of transportation modelling. The main purpose of which are traffic control, reporting of accidents, avoiding traffic jams due to knowledge of traffic flow and subsequent transportation planning. A number of solutions exist — both model-driven and data driven had proven to be successful in capturing the dynamics of traffic flow. Nevertheless, most space-time models suffer from high mathematical complexity and low efficiency. Artificial Neural Networks, one of the prominent datadriven approaches, show promising performance in modelling the complexity of traffic flow. We present a neural network architecture for traffic flow prediction on a real-world road network graph. The model is based on the combination of a recurrent neural network and graph convolutional neural network. Where a recurrent neural network is used to model temporal dependencies, and a convolutional neural network is responsible for extracting spatial features from traffic. To make multiple few steps ahead predictions, the encoder-decoder architecture is used, which allows to reduce noise propagation due to inexact predictions. To model the complexity of traffic flow, we employ multilayered architecture. Deeper neural networks are more difficult to train. To speed up the training process, we use skip-connections between each layer, so that each layer teaches only the residual function with respect to the previous layer outputs. The resulting neural network was trained on raw data from traffic flow detectors from the US highway system with a resolution of 5 minutes. 3 metrics: mean absolute error, mean relative error, mean-square error were used to estimate the quality of the prediction. It was found that for all metrics the proposed model achieved lower prediction error than previously published models, such as Vector Auto Regression, LSTM and Graph Convolution GRU.

The reconstruction of charged particle trajectories in tracking detectors is a key problem in the analysis of experimental data for high energy and nuclear physics.

The amount of data in modern experiments is so large that classical tracking methods such as Kalman filter can not process them fast enough. To solve this problem, we have developed two neural network algorithms of track recognition, based on deep learning architectures, for local (track by track) and global (all tracks in an event) tracking in the GEM tracker of the BM@N experiment at JINR (Dubna). The advantage of deep neural networks is the ability to detect hidden nonlinear dependencies in data and the capability of parallel execution of underlying linear algebra operations.

In this work we generalize these algorithms to the cylindrical GEM inner tracker of BESIII experiment. The neural network model RDGraphNet for global track finding, based on the reverse directed graph, has been successfully adapted. After training on Monte Carlo data, testing showed encouraging results: recall of 98% and precision of 86% for track finding.

The local neural network model TrackNETv2 was also adapted to BESIII CGEM successfully. Since the tracker has only three detecting layers, an additional neuro-classifier to filter out false tracks have been introduced. Preliminary tests demonstrated the recall value at the first stage of 99%. After applying the neuro-classifier, the precision was 77% with a slight decrease of the recall to 94%. This result can be improved after the further model optimization.

We suggested some approaches for solving image processing tasks in the decision support system (DSS) of the Center for Automated Recording of Administrative Offenses of the Road Traffic (CARAO). The main task of this system is to assist the operator in obtaining accurate information about the vehicle registration plate and the vehicle brand/model based on images obtained from the photo and video recording systems. We suggested the approach for vehicle registration plate recognition and brand/model classification on the images based on modern neural network models. LPRNet neural network model supplemented by Spatial Transformer Layer was used to recognize the vehicle registration plate. The ResNeXt-101-32x8d neural network model was used to classify for vehicle brand/model. We suggested the approach to construct the training set for the neural network of vehicle registration plate recognition. The approach is based on computer vision methods and machine learning algorithms. The SIFT algorithm was used to detect and describe local features on images with the vehicle registration plate. DBSCAN clustering was used to detect and delete outliers in such local features. The accuracy of vehicle registration plate recognition was 96% on the testing set. We suggested the approach to improve the efficiency of using the ResNeXt-101-32x8d model at additional training and classification stages. The approach is based on the new architecture of convolutional neural networks with “freezing” weight coefficients of convolutional layers, an additional convolutional layer for parallelizing the classification process, and a set of binary classifiers at the output. This approach significantly reduced the time of additional training of neural network when new vehicle brand/model classification was needed. The final accuracy of vehicle brand/model classification was 99% on the testing set. The proposed approaches were tested and implemented in the DSS of the CARAO of the Republic of Tatarstan.

The numerical methods used to simulate the evolution of hydrodynamic systems require the considerable use of computational resources thus limiting the number of possible simulations. The data-driven simulation technique is one promising approach to the development of heuristic models, which may speed up the study of such models. In this approach, machine learning methods are used to tune the weights of an artificial neural network that predicts the state of a physical system at a given point in time based on initial conditions. This article describes an original neural network architecture and a novel multi-stage training procedure which create a heuristic model of a two-phase flow in a heterogeneous porous medium. The neural network-based model predicts the states of the grid cells at an arbitrary timestep (within the known constraints), taking in only the initial conditions: the properties of the heterogeneous permeability of the medium and the location of sources and sinks. The proposed model requires orders of magnitude less processor time in comparison with the classical numerical method, which served as a criterion for evaluating the effectiveness of the trained model. The proposed architecture includes a number of subnets trained in various combinations on several datasets. The techniques of adversarial training and weight transfer are utilized.