Responsible AI for Networked Systems

I will describe a responsible methodology for applying AI to networked systems that is minimally disruptive, synergistic with human solutions, and safe. First, I will develop a paradigm that combines reinforcement learning with the ability to ask counterfactual (“what if”) questions about a decision-making system and show how to use this to exploit the natural information emitted by these systems. Then, I will describe an abstraction called a “safeguard” that protects an AI system from violating a safety specification, while allowing the system and the safeguard to co-evolve. We will apply this methodology to several infrastructure systems in our Azure cloud and edge.