Bayesmix: sharp loss bounds Marcus Hutter (on Schmidhubers SNF grant; IJFCS / ECML / ICML 2001)
Predict k-th symbol: loss in [0,1].
Optimal: minimize µ-expected loss
Instead: minimize M-expected loss
Total loss difference for first n symbols:
If µ deterministic then soon no more errors!
Back to J. Schmidhuber's Theory of Everything page