Journal of Network and Systems Management Information Systems
Orchestrating Artificial Intelligence Functions
Abstract
Abstract The convergence of cloud computing and mobile networks is enabling a new generation of AI-enabled applications composed of multiple interdependent processing stages distributed across the edge-cloud continuum. In this paper, we introduce the concept of Artificial Intelligence Functions (AIFs) as modular, orchestratable building blocks of such applications. We further propose a functional split between an ML Data Plane, carrying data flows between AIFs, and an ML Control Plane, carrying model parameter exchanges such as those arising in centralized retraining, model versioning, or federated learning.. Users specify their application requirements as an Artificial Intelligence Forwarding Graph (AIFG), a directed acyclic graph (DAG) capturing computational, hardware acceleration, latency, and bandwidth requirements of each AIF and its interconnections. We formalize the AIFG embedding problem as an Integer Linear Programming (ILP) model that minimizes placement cost while satisfying resource and latency constraints, and we propose a scalable greedy heuristic for larger problem instances. We evaluate both approaches using CPLEX across substrate topologies of increasing size and varying AIFG complexities. Results show that the heuristic achieves near-optimal embedding cost while reducing computation time by several orders of magnitude, making it suitable for practical deployment scenarios. We further validate the practical feasibility of the proposed framework through a cloud-native software prototype.