3 ms·
Thanks for the clarification. So I take it that means I should read the following: The key feature of reflective oracles is that they avoid diagonalization an
by kbenson 10y ago
Thanks for the clarification. So I take it that means I should read the following:
The key feature of reflective oracles is that they avoid diagonalization and paradoxes by randomizing in the relevant cases.2 This allows agents with access to a reflective oracle to consistently reason about the behavior of arbitrary agents that also have access to a reflective oracle, which in turn makes it possible to model agents that converge to Nash equilibria by their own faculties (rather than by fiat or assumption).
as not that the agents will converge on a Nash equilibria, but that now it's possible to model agents that do converge on a Nash equilibria, which previously we could not (at at least not definitively)? That's what it seems to say outright, which I missed previously, I just want to make sure I'm not approaching it from the wrong context.
- danbruc 10y agoDon't quote me on that, I have really nothing to do with game theory, but as I understand it the paper it shows that reflectiveoracle-computable policies are optimal in the discussed setup (Theorem 25) and that they will yield a Nash equilibrium if all policies are asymptotically optimal (Theorem 28) which is possible because a limit computable reflective oracle exists (Theorem 6) and Thompson sampling is asymptotically optimal. So achieving a Nash equilibrium still requires that all agents play along, they have to have asymptotically optimal policies, if other agents show erratic behavior you are unable to make sense of, you can not respond optimally.
- AstralStorm 10y agoYou can still respond optimally in a local sense, that is derive a dominant strategy. It will take longer than an agent that is not reflective though, at least by one sampling step.