While machine learning and artificial intelligence technologies are now advanced enough to outperform humans in variety of tasks, how we make decisions with models varies by practitioner. Reinforcement learning is promising but it is limited to adversarial settings; or, in vernacular, situations where decisions directly impact the environment. Without figuring out how AI systems can make good decisions in environments they cannot influence, we may forever be stuck in a limbo of pattern recognition, prediction, and analytics. What if we can develop a “theory” of AI decision making? Can we view different decision making situations as a set of engineering systems? Can we define key components of an AI decision maker? Answering such questions would enable us to design AI systems in a modular fashion much like how we design many industrial goods like cars. We may even be able to develop industry standards and manuals on how to design AI decision makers. Using actual use cases and other potential real-world applications in both financial and non-financial settings as examples, a systems view on decision making AI systems is proposed. Furthermore, ways to design and build such systems are explored.
1) AI systems are far more valuable making decisions than making simple predictions and pattern recognitions
2) We need a “theory of design” for AI decision making systems: For AI to become a trusted part of decision making both in and out of industry, we need to understand and generalize components of AI systems and define what it means to be “robust”.
3) This starts with looking at actual use cases, identifying similarities, and quantifying key parameters – so we may use standard design techniques to design AI systems.
Kevin is a data scientist with strong interest in an interdisciplinary approach that combines artificial intelligence, operations research, systems engineering, economics, and quantitative finance. He is a member of Nasdaq’s Machine Intelligence Lab, a team dedicated to using AI to improve capital markets. At Nasdaq, he has worked on projects that cover topics such as alternative data, capital market operations, financial surveillance, and portfolio management. His most recent interest is in developing a procedure for designing robust and fail-proof decision-making AI systems. Kevin holds a Bachelor’s degree in Computer Science from Washington University in St. Louis.