Аdaptive method for service placement in fog computing systems based on graph attention and multi-agent reinforcement learning

Abstract

The paper presents the Adaptive Fog Architecture with Integrated Artificial Intellegence method for constructing an optimal fog computing system architecture. The method implements a four-component decision-making pipeline: a trained graph attention layer for structural encoding of network topology; a multi-agent controller of the Multi-Agent Deep Deterministic Policy Gradient type with shared Actor-network weights and a mean-field critic for scalability across a variable number of nodes; an adaptive orchestrator based on a stochastic Boltzmann auction; and a feedback module that trains the model on actually realized subtask execution outcomes rather than a priori estimates. The method is implemented and evaluated in a discrete-event simulation environment (SimPy) featuring realistic service queues, channel bandwidth contention, chained subtask dependencies, and a stochastic node failure process. A comparative evaluation across eight methods (the presented method and seven baselines: Particle Swarm Optimization, Genetic Algorithm, Ant Colony Optimization, Machine Learning + Particle Swarm Optimization, Multi-Agent Reinforcement Learning, Federated Learning, and Multi-Agent Reinforcement Learning + Graph Neural Networks + Federated Learning) in five scenarios under baseline and stress loads showed that the presented method achieves the best load-balancing performance among all eight methods in four of the five baseline scenarios, along with competitive, though not best, average latency. It was established that the decision-making paradigm (periodic batch scheduling versus immediate online response) is the dominant factor distinguishing method classes by latency, exceeding in significance the choice of a specific optimization algorithm within a single architecture. An architectural trade-off was identified: auction stochasticity, beneficial for load balancing under normal conditions, reduces the deadline compliance rate under extreme load. Independent holdout validation of the initially proposed method-class classifier did not confirm the hypothesis of a determining role of load variability, substantiating the need to reformulate it based on architectural parameters of the system rather than statistical characteristics of the input stream.