Positive feedback loops will become essential for superintelligence, and non-cooperative games can produce one of the strongest intelligence-amplifying feedback loop when they (1) exert "intelligence pressure" on players and (2) the players learn.
I believe we've already seen the early signs of this potential in GANs and self-play. However, fundamental questions remain largely open: how and when multiple neural nets would actually converge to their equilibrium, and what'd be the precise nature of such an equilibrium?
In smooth games, each player optimizes its own objective using gradients. The challenge is that competing objectives often induce non-convergent cycles. In fact, Nash equilibria and stationary points are computationally intractable in general. So the theory typically splits into two distinct categories: local analysis under general setup and global analysis under structural assumptions. My AISTATS 2022 paper studied the former and my latest COLT 2026 paper studies the latter. Now I'm working on a a fundamental question: "what'd be the minimal structure of tractable smooth games?".
In real-world applications, we seldom enjoy gradient feedback, and we only have access to stochastic rewards. Markov games naturally extend to this setting, and classical zero-sum Markov games are now relatively well-established. However, Markov games with general-sum, function approximation, and non-additive rewards are still in their infancy. I'm intrigued by the question: "what'd be the minimal structure of tractable Markov games?".
In the world of minimization, deep learning now has a plausible hypothetical explanation: neural nets admit multiple solutions, but stochastic gradient methods prefer particularly desirable ones. I believe an analogous explanation would be crucial to understand games. I'm working on the question: "given a continuum of equilibria or limit cycles, which one would stochastic gradient dynamics prefer?"
I'm actively looking for a Ph.D. opportunity at the intersection of non-cooperative games and deep learning. I'd be happy to discuss research with anyone interested. Please feel free to reach out to me via email: junsoo.ha.contact@gmail.com