TARK-2025
Program

The talk will begin with a discussion of unexpected ways in which the advance of scientific knowledge can backfire on human welfare in game-theoretic environments. It will then go on to analyze other kinds of interface between knowledge and well-being, in the area of political economy and governance, which sheds interesting light on the rise of authoritarianism and the erosion of democracy. The presentation will end with tentative comments on how individuals in strategic environments can think about creating a priori behavioral rules and laws to avoid some of the pitfalls of knowledge. Inevitably, there will be more open questions than answers.

Classification algorithms are increasingly used in high-stakes domains such as credit, employment, healthcare, housing, law enforcement, and national security. These systems do more than assign labels--they influence behavior by shaping incentives and expectations. In this lecture, I argue that when classification affects individual behavior, conventional statistical notions of algorithmic fairness can be misleading or incomplete. I introduce a social scientific framework for modeling behavioral responses to classification and analyzing how classification rules shape outcomes across groups. This framework yields new fairness criteria that account for endogenous behavior, individual welfare, and explicability. I'll draw tight connections between statistical, behavioral, welfare, and transparency-based fairness criteria, and identify conditions under which they can align. The results underscore the need to move beyond error-based metrics to evaluate fairness in systems that interact with, and shape, human behavior.

In the “Savage” model of decision-making under uncertainty, there is a space S of possible “states of nature” (representing the source of uncertainty), and a space X of possible “outcomes”. Each course of action (“act”) is represented as a function from S to X. A rational agent has a utility function on X and (probabilistic) “beliefs” about S, and evaluates each act according to its expected utility.

However, a single agent might encounter many different sources of uncertainty and many different menus of outcomes, which could be combined together into many different decision problems. Furthermore, these different uncertainty sources (or outcome menus) might be related to one another in various ways:

1. There may be analogies between different uncertainty sources (or different outcome menus).
2. At different times, the agent may also have different levels of awareness, or access to different informational resources. Different uncertainty sources (or outcome menus) could represent the agent's subjective perception of the same objective decision problem with different levels of awareness or information.
3. Some uncertainty sources (or outcome menus) might exhibit internal symmetries.

Furthermore, in some situations, the state spaces and outcome spaces have additional mathematical structure (e.g. a topology or differentiable structure), and feasible acts must respect this structure (i.e. they must be continuous or differentiable functions). In other situations, the agent might only have access to linguistic descriptions of background conditions, actions, and their consequences (e.g. encoded in a Boolean algebra or some other algebraic structure). Finally, there may be cases where the agent is only cognizant of a set of abstract “acts”, and is unable to specify explicit state spaces and outcome spaces.

I introduce a modelling framework that addresses all of these issues. I then define and axiomatically characterize a subjective expected utility representation that is “global” in two senses. First: it posits probabilistic beliefs for all uncertainty sources and utility functions over all outcome menus, which simultaneously rationalize the agent’s preferences across all possible decision problems, and which are consistent with the aforementioned analogies, symmetries, and awareness levels. Second: it applies in many different mathematical environments (i.e., different categories), making it unnecessary to develop a separate theory for each one.

Papers:

• Global Subjective Expected Utility Representations (working paper)
  (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5039921)
• Subjective Expected Utility on Orthomodular Lattices (to appear in Philosophical Transactions of the Royal Society A, 2025)
  (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5048137)

Since the advent of AI, games have served as progress benchmarks, and most real-world settings are imperfect-information games. Meanwhile, imperfect-information variants of chess have existed for over a century, present extreme challenges, and have been the focus of significant AI research. Beyond calculation needed in regular chess, they require reasoning about information gathering, the opponent’s knowledge, signaling, bluffing, etc. The most popular variant, Fog of War (FoW) chess (aka. dark chess) is a recognized challenge problem in AI after superhuman performance was reached in no-limit Texas hold’em poker. We present Obscuro, the first superhuman AI for FoW chess. It introduces advances to search in imperfect-information games, enabling strong, scalable reasoning. Most prior search techniques - such as those used to achieve superhuman play in no-limit Texas hold’em - require the construction of the “common knowledge set” as a first step, making them unusable for games with this much imperfect information. Experiments against the prior state-of-the-art AI and human players - including the world’s best - show that Obscuro is significantly stronger. FoW chess is now the largest (by amount of imperfect information) turn-based game in which superhuman performance has been achieved and the largest game in which imperfect-information search has been successfully applied.

This is joint work with my PhD student Brian Hu Zhang.