DEEP REINFORCEMENT LEARNING FOR REAL-TIME ADAPTIVE USER WEB-INTERFACE CUSTOMIZATION

Abstract

Purpose. The purpose of the work is to develop adaptive web interface design using machine learning with reinforcement methods. Methodology. The research methodology was based on the application of deep reinforcement learning for adaptive customization of the web interface in real time. User interaction with the interface was modeled as a decision-making process. The Markov Decision Process (MDP) technology was used, where the policy determines changes to interface elements based on a reward function that integrates user engagement metrics (click rate, idle time) and compliance with WCAG accessibility standards. Maximum likelihood estimators and gradient optimization were used to determine effective interaction patterns. The analysis of time dependencies of interactions was carried out using the LSTA and LSTM modules, which focus on short-term and long-term user preferences, using convolutional layers and an attention mechanism to extract relevant f eatures. Results. To analyze user behavior patterns, including navigation paths, input methods, and customization options, the framework uses maximum likelihood estimation and gradient-based optimization. This process models user preferences and interactions to efficiently optimize user interface settings. In addition, the architecture integrates a long-short-term memory (LSTM) module to capture temporal dependencies in user interactions. By processing data through convolutional gating layers, the network learns both instantaneous and extended user preferences. Temporal aggregation is further enhanced by adaptive convolutional kernels and attention mechanisms that prioritize critical areas of the interface based on user interaction. The adaptability of the system is improved by incorporating fuzzy logic into the gating mechanisms, where fuzzy rules model the complex dependencies between factors that affect user interaction. This integration facilitates robust decision-making under different conditions, ensuring stability and preventing overfitting. The LSTM module uses convolutional operations on merged cells and hidden states to extract vital features, solve the problem of gradient vanishing, and improve memory retention. Actions generated by the model dynamically update user interface elements, contributing to a seamless and personalized user experience. The backend implementation, built on FastAPI, processes user interaction data in real time, calculates rewards, and predicts actions to optimize user interface settings. The server uses TensorFlow-based models to analyze user metrics and accessibility data, providing useful information for interface customization. The architecture provides efficient communication between the backend and the frontend, allowing for real-time user interface modification. The scientific novelty of the article lies in the development of an adaptive web interface model that uses deep reinforcement learning to personalize user interactions in real time, taking into account temporal dependencies and accessibility metrics. The practical relevance is determined by the possibility of applying the proposed approach to increase user engagement and ensure inclusiveness of web resources, which is relevant for various industries, including e-commerce and educational platforms.