High-Confidence Predictions under Adversarial Uncertainty

TL;DR

This paper develops algorithms for high-confidence predictions of unseen bits in an infinite binary sequence under limited assumptions, including weak sparsity and automaton-based constraints, with applications to forecasting and sequential prediction tasks.

Contribution

It introduces a randomized prediction algorithm for eps-weakly sparse sequences and extends to broad automaton-based assumptions, advancing sequential prediction under uncertainty.

Findings

Predicts a zero with success probability close to 1 - eps for eps-weakly sparse sequences.

Provides a forecasting algorithm that predicts the fraction of ones within an epsilon margin with high confidence.

Extends to automaton-based assumptions, broadening the applicability of sequential prediction methods.

Abstract

We study the setting in which the bits of an unknown infinite binary sequence x are revealed sequentially to an observer. We show that very limited assumptions about x allow one to make successful predictions about unseen bits of x. First, we study the problem of successfully predicting a single 0 from among the bits of x. In our model we have only one chance to make a prediction, but may do so at a time of our choosing. We describe and motivate this as the problem of a frog who wants to cross a road safely. Letting N_t denote the number of 1s among the first t bits of x, we say that x is "eps-weakly sparse" if lim inf (N_t/t) <= eps. Our main result is a randomized algorithm that, given any eps-weakly sparse sequence x, predicts a 0 of x with success probability as close as desired to 1 - \eps. Thus we can perform this task with essentially the same success probability as under the much stronger assumption that each bit of x takes the value 1 independently with probability eps. We apply this result to show how to successfully predict a bit (0 or 1) under a broad class of possible assumptions on the sequence x. The assumptions are stated in terms of the behavior of a finite automaton M reading the bits of x. We also propose and solve a variant of the well-studied "ignorant forecasting" problem. For every eps > 0, we give a randomized forecasting algorithm S_eps that, given sequential access to a binary sequence x, makes a prediction of the form: "A p fraction of the next N bits will be 1s." (The algorithm gets to choose p, N, and the time of the prediction.) For any fixed sequence x, the forecast fraction p is accurate to within +-eps with probability 1 - eps.