Finite-Memory Prediction as Well as the Empirical Mean

TL;DR

This paper investigates the limits of finite-state machines in predicting continuous sequences, comparing different machine classes and introducing new models to minimize prediction regret relative to the empirical mean.

Contribution

It analyzes the tradeoff between the number of states and regret, introduces the DTM and EDM machines, and proposes an enhanced machine that outperforms existing models.

Findings

Optimal small-state machine is DTM.

EDM achieves vanishing regret with many states.

Enhanced EDM outperforms EDM for all state counts.

Abstract

The problem of universally predicting an individual continuous sequence using a deterministic finite-state machine (FSM) is considered. The empirical mean is used as a reference as it is the constant that fits a given sequence within a minimal square error. With this reference, a reasonable prediction performance is the regret, namely the excess square-error over the reference loss, the empirical variance. The paper analyzes the tradeoff between the number of states of the universal FSM and the attainable regret. It first studies the case of a small number of states. A class of machines, denoted Degenerated Tracking Memory (DTM), is defined and the optimal machine in this class is shown to be the optimal among all machines for small enough number of states. Unfortunately, DTM machines become suboptimal as the number of available states increases. Next, the Exponential Decaying Memory (EDM) machine, previously used for predicting binary sequences, is considered. While this machine has poorer performance for small number of states, it achieves a vanishing regret for large number of states. Following that, an asymptotic lower bound of O(k^{-2/3}) on the achievable regret of any k-state machine is derived. This bound is attained asymptotically by the EDM machine. Furthermore, a new machine, denoted the Enhanced Exponential Decaying Memory machine, is shown to outperform the EDM machine for any number of states.