Fellows' Spotlight

Advanced AI systems may have many novel ways of credibly committing to and enforcing agreements in strategic interactions. But even when players can make arbitrary joint commitments, it may be unclear whether a particular agreement actually constitutes a Pareto improvement. Players might not know or might not agree on how the game would be played by default, and strategic posturing might prevent them from revealing their true beliefs (cf. equilibrium selection).

In this talk, Nathaniel will discuss his work on Safe Pareto Improvements (SPIs), an approach which circumvents these obstacles. Rather than agreeing on particular outcomes, we seek commitments which leave the game strategically equivalent ('isomorphic') to the original, but with payoffs that constitute a Pareto improvement. Nathaniel will give geometric characterisations for when and how such SPIs exist, describe ongoing work on affordances which enable them in larger classes of games, and conclude by discussing some open questions for future SPI research.