Multi-agent method to improving adaptive real-time management of computing resources

The paper presents the development of a multi-agent method aimed at meeting the growing demand for computing resources by enhancing the adaptability and efficiency of real-time management. In practical scenarios, it is essential to enable prompt and flexible adaptive adjustments to the task execution schedule in order to improve overall resource utilization. The proposed multi-agent resource management method is based on a previously developed "network of needs and capabilities" model, which enables smooth, adaptive modifications to the execution schedule. This process involves a sequence of atomic stepwise changes to the resource allocation plan, including local task shifts within a single computing resource, as well as the displacement and redistribution of tasks across multiple resources. Each task agent calculates an optimal “patch” to maximize the global efficiency of the system, accounting for the satisfaction functions of all tasks affected by the change. A key innovation is the introduction of a collective decision-making mechanism based on the computation and coordination of these patches. This allows for dynamic optimization of the schedule without requiring full rescheduling or transitioning to a fully decentralized solution, which would eliminate the shared data environment of the agent system. Experimental results demonstrate that the proposed method increases system efficiency by 25–30% compared to non-adaptive control approaches, which lack the ability to selectively revise agent interactions or reallocate tasks among resources. The method also enhances the scalability and fault tolerance of the system, expanding its applicability to a broad range of dynamic resource allocation problems in computing, manufacturing, and logistics.