New Research Explores Machine Learning Safety Without Countless Trials

Artificial intelligence (AI) technology has long taken inspiration from human learning processes. However, as AI-based systems like self-driving cars and power grids move towards critical autonomous applications, the risk to human safety has grown. A recent study challenges the belief that an unlimited number of trials is necessary to produce safe AI machines. The research introduces a new approach based on balancing safety with optimality, encountering hazardous situations and rapidly identifying unsafe acts. By using the mathematical framework known as the Markov decision process, the team found a way to accelerate and manage tradeoffs between optimality, detection time and exposure to unsafe events. The research has implications for robotics, autonomous systems and AI safety as a whole.