Raja Panjwani

Artificial intelligence systems are increasingly used to assist humans in sequential decision-making tasks. A common approach is to recommend actions that are optimal according to a strong model of the environment. Such recommendations, however, implicitly assume that the decision maker will continue to act optimally afterward. Human decision makers often deviate from optimal play, meaning that actions that are optimal in isolation may lead to states that are difficult for humans to navigate successfully. This raises a fundamental question: how should AI systems intervene when assisting suboptimal decision makers in sequential environments? In this work, we study value-aware interventions that account for how humans are likely to behave after receiving assistance. Our approach builds on a basic principle from reinforcement learning: when a policy is suboptimal, discrepancies arise between the actions taken by the policy and those that maximize the expected value of future outcomes. We show how such discrepancies can be used to identify opportunities where interventions improve downstream performance. To operationalize this idea, we develop a data-driven framework that learns models of human decision-making from large datasets and uses these models to guide intervention decisions. We study these ideas in the domain of chess, which provides large-scale human gameplay data and strong AI systems for evaluating intervention strategies.

The use of linear regression is ubiquitous across the social and behavioral sciences, and yet researchers rarely hold that the variables in their target systems are in fact linearly related. This raises the question of how to interpret linear regression coefficients when there is ‘functional misspecification’ and the target system exhibits nonlinearity. Here, recent methodological discussions among practitioners have mirrored philosophical debates over scientific realism. In this paper, I frame the prevailing views in terms of their stance on the scientific realism debate. I then present a novel realist interpretation of linear regression—the ‘secant interpretation’—based on a property I derive about the manner in which linear regression coefficients represent properties of non-linear systems of interest.

Interpreting judicial decision-making in terms of a realistic optimization problem has been a longstanding challenge for economic theories of adjudication. This article develops a theory of value for a fundamental decision facing an appellate court: the choice of whether to follow, distinguish, or overrule precedent. Extending the framework of reason-based choice into a dynamic model of jurisprudence, the theory illustrates distinctive judicial phenomena such as the practice of ‘stealth overruling’, and also sheds new light on judicial minimalism, legal completeness, and the constraint of precedent.

Scientists routinely draw inferences about target systems by deriving them from highly idealized models, and their derivations tend to explicitly rely on assumptions which are clearly false of their targets. Such practices would appear to be unsound, and their justification presents a challenge in model epistemology. To meet this challenge, we develop a method, ‘Soundness Analysis’, for decomposing an idealized model into its true and false aspects in order to determine whether a model’s true aspects can serve as premises in a sound argument for the model’s conclusion. We illustrate the method on the Hoff-Stiglitz model from economics, and we argue that Soundness Analysis has epistemic advantages over prevailing de-idealization strategies.