An adversarial risk analysis framework for software release decision support.

Abstract

Recent artificial intelligence (AI) risk management frameworks and regulations place stringent quality constraints on AI systems to be deployed in an increasingly competitive environment. Thus, from a software engineering point of view, a major issue is deciding when to release an AI system to the market. This problem is complex due to, among other features, the uncertainty surrounding the AI system's reliability and safety as reflected through its faults, the various cost items involved, and the presence of competitors. A novel general adversarial risk analysis framework with multiple agents of two types (producers and buyers) is proposed to support an AI system developer in deciding when to release a product. The implementation of the proposed framework is illustrated with an example and extensions to cases with multiple producers and multiple buyers are discussed.