Robust Decision Heuristics for Natural and Artificial Intelligence

Robust Decision Heuristics for Natural and Artificial Intelligence

September 2025 - August 2028

This long-term group seeks to bridge the study of decision-making under uncertainty in natural systems and the design of robust AI systems. Here, "natural systems" is primarily a reference to humans but we will also consider insects because they have simpler decision processes. Specifically, we will:

1. explore the design of algorithms for robust and transparent AI systems by drawing on models of decision processes under uncertainty in natural systems; and
2. explore the design of AI systems driven by decision heuristics to gain new understanding about decision-making in natural systems.

 

Decision heuristics build on Herb Simon's definition of Bounded Rationality. They use adaptive satisficing to make rational decisions "in the wild", i.e., under open world uncertainty, when the space of possible scenarios is not known in advance. They comprise simple methods such as tallying, sequential search, and fast and frugal trees, and an algorithmic approach to identify heuristics-methods that match the characteristics of the domain and tasks at hand. They have demonstrated the ability to make decisions more quickly, frugally, and accurately than more complex methods in different application domains, and to automatically provide process-level explanations of the decisions. However, they have not been fully leveraged in the design of robust AI systems or in enhancing our understanding of decision-making in natural systems. Instead, most computationally methods considered state of the art in AI (and related disciplines such as psychology, social science, and neuroscience) are based on increasingly complex mathematical models with a narrow set of representations and update processes; they are also resource hungry, opaque, and are not designed to be robust in non-stationary environments. In addition, there is limited work in the design and use of decision heuristics to understand and develop psychologically valid theories of decision-making in natural systems.

 

Our long-term group pursues a different philosophy for addressing these knowledge gaps and the underlying scientific problems. We will design and use decision heuristics on their own and in combination with existing formalisms to deepen our understanding of decision-making in natural systems and to enable the widespread design of robust and transparent AI systems.