AI decision-making is starting to play a critical role in intelligent decision making in complex networks such IoT, cyber-physical systems, or intelligent edge. However, existing AI algorithms are often not practically feasible in these networks, as they hinge on excessive use of cyber resources, e.g., in terms of energy, communication, and local computation. The goal of this project is to advance the systematic design and algorithmic foundations of resource efficient ML in complex networks. We focus on multi-agent reinforcement learning algorithms, where agents learn by experience how to use resources as efficiently as possible while performing their AI tasks. We plan to use the SNIC computational infrastructure to test our novel algorithm designs for resource-efficient AI.