Recency-weighted Markovian inference

TL;DR

This paper introduces a recency-weighted Markovian inference model that uses a decaying mixture over past states, along with an efficient sampling algorithm for high-order models.

Contribution

It proposes a novel MLSS model with recency weighting and a simple sampling algorithm that maintains fixed computational costs.

Findings

Efficient approximation of high-order MLSS models.

Fixed time and memory costs for inference.

Decaying mixture over multiple past states.

Abstract

We describe a Markov latent state space (MLSS) model, where the latent state distribution is a decaying mixture over multiple past states. We present a simple sampling algorithm that allows to approximate such high-order MLSS with fixed time and memory costs.