A Multicriteria Method for Optimizing it Service Management of a Virtual Provider Based on Deep Reinforcement Learning

Abstract

This article addresses the problem of optimizing IT service management for the B2B segment under conditions of dynamic workloads and the probabilistic unreliability of service operators. The architecture of a Virtual Service Provider (VSP) management system is proposed, which automates the service processes for Corporate Customers. The core of the system is a hybrid translation module that mathematically transforms abstract business intents and client context into a deterministic Service-Resource Model with specified technical, financial, and time constraints. To efficiently orchestrate the generated tasks, a multicriteria optimization algorithm, PPO-VSP, based on deep reinforcement learning (Actor-Critic architecture) was developed. The implementation of a reputation assessment module allowed the system to identify unreliable service providers and avoid overloading them. Experimental studies using simulation modeling confirmed the stable mathematical convergence of the algorithm. The trained optimization agent ensured compliance with service level agreements (the Quality of Experience metric) at a level of 95.4% under limited resource conditions. The generation time of the decomposition matrix averaged 18 ms, guaranteeing rapid management decision-making during the operation of high-load infrastructures without system downtime.