A temporally abstracted Viterbi algorithm
This work addresses computational efficiency in sequence decoding for applications like signal processing or bioinformatics, representing an incremental advance over prior coarse-to-fine methods.
The paper tackles the problem of speeding up the Viterbi algorithm by applying temporal abstraction, resulting in improvements of several orders of magnitude over the standard Viterbi algorithm and significant speedups over coarse-to-fine dynamic programming for problems with state variables evolving at widely differing rates.
Hierarchical problem abstraction, when applicable, may offer exponential reductions in computational complexity. Previous work on coarse-to-fine dynamic programming (CFDP) has demonstrated this possibility using state abstraction to speed up the Viterbi algorithm. In this paper, we show how to apply temporal abstraction to the Viterbi problem. Our algorithm uses bounds derived from analysis of coarse timescales to prune large parts of the state trellis at finer timescales. We demonstrate improvements of several orders of magnitude over the standard Viterbi algorithm, as well as significant speedups over CFDP, for problems whose state variables evolve at widely differing rates.