A method for adaptive replication and data migration in heterogeneous networks with a dynamic structure

Abstract

The relevance of this research lies in resolving the conflict between ensuring high availability and integrity of information and the resource constraints (power consumption, memory capacity, bandwidth) of mobile terminals in a heterogeneous network with a stochastically varying topology. Traditional approaches based on distributed hash tables, static replication, and NoSQL databases are limited by high communication overhead, degradation of routing during frequent connection drops, and an inability to proactively predict environmental dynamics, which necessitates the development of effective mathematical and software solutions for storing structured data. The goal of this work is to improve the availability, semantic integrity, and energy efficiency of structured data storage in heterogeneous networks with a dynamic structure by developing a hybrid model of adaptive replication controlled by a Deep Q-Network. To achieve this goal, the following tasks have been set: to formalize the access cost objective function, taking into account the resource parameters of mobile nodes; to develop a method for adaptive data replication and migration in the form of a machine learning pipeline optimized for deployment on edge devices; to develop the topology of a Deep Q-Network for adaptive control of the placement, migration, and competitive replacement of replicas in a stochastic topological environment. The research methods are based on the theory of Markov decision processes, reinforcement learning algorithms, and methods of mathematical statistics for analyzing data drift. The results of this work include the formalization of a method for adaptive replication of structured data guided by a neural network model based on criteria of expected channel lifetime and weighted demand, as well as the development of a machine learning pipeline using deep reinforcement learning algorithms optimized for deployment on edge devices. An experimental study of the developed solutions on a simulation platform using spatiotemporal real-world mobility datasets demonstrated that the proposed method maintains a query delivery success rate of eighty-three point nine percent at node speeds of up to one hundred kilometers per hour (compared to forty-six point nine percent for static replication and forty-one point one percent for Bamboo DHT), reduces the average search time by fifty-eight percent (from two hundred seven milliseconds to eighty-seven milliseconds), reduces communication energy consumption by thirty percent, and decreases the number of storage rewriting cycles by fifty percent. The findings confirm the high effectiveness of using DRL orchestration to ensure a controlled inconsistency regime under conditions of spatial isolation of network segments. The scientific novelty of the obtained results lies in the development of a method for adaptive replication of structured data in heterogeneous networks with a dynamic structure by continuously determining the replication coefficient and the optimal nodes for placing data copies based on the prediction of node trajectories and the level of their hardware and energy resources. The practical significance lies in the adaptation of an optimized machine learning pipeline, which enables the direct implementation of the model in autonomous transportation, environmental monitoring sensor networks, and emergency response systems.