← All problems

The partition-based updating conjecture for minimax decision rules

19 years old · traced to

Let P{\cal P} be an arbitrary set of distributions on X×Y{\cal X}\times{\cal Y}, let LL be an arbitrary loss function on Y×A{\cal Y}\times{\cal A}, and let δ\delta be the minimax decision rule relative to LL and P{\cal P}. For a function φ:X→R\varphi:{\cal X}\to\mathbb{R}, write P(Y)(⋅∣φ(X)=φ(x)){\cal P}^{(Y)}(\cdot\mid\varphi(X)=\varphi(x)) for the set of conditional distributions of YY induced by distributions in P{\cal P} at the level φ(x)\varphi(x). The minimax updating conjecture. There exists a function φ:X→R\varphi:{\cal X}\to\mathbb{R} such that

δ(x)=arg⁡min⁡a∈Amax⁡Pr⁡∈P(Y)(⋅∣φ(X)=φ(x))EPr⁡[L(Y,a)].\delta(x)=\arg\min_{a\in{\cal A}}\max_{\Pr\in{\cal P}^{(Y)}(\cdot\mid\varphi(X)=\varphi(x))}E_{\Pr}[L(Y,a)].

Thus, the minimax-optimal decision rule would be a Π\Pi-based rule, with the probability update rule Π\Pi given by φ\varphi-conditioning at some level of φ\varphi. The conjecture is false, as demonstrated by a counterexample in the paper.

References
  1. Peter D. Grunwald and Joseph Y. Halpern, “A Game-Theoretic Analysis of Updating Sets of Probabilities”, arXiv:0711.3235 (2007).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.