04Classifiers and decoders

Hidden Markov models (hmmlearn)

The generative sequence model: hidden states, noisy observations, exact inference. Forward-backward gives per-frame posteriors; Viterbi gives the single best path.

  • Sequence
  • Classical
  • Labels optional
  • BSD-3-Clause
Forward-backward combines past, present and future evidence into a confidence for each state.
Modality
Per-frame features or probabilities
Output
State posteriors and the best state path
Evidence
Rabiner's tutorial (1989); long-standing library
Code and weights
hmmlearn, in limited-maintenance mode; no football parameters
Licence
BSD-3-Clause

A hidden Markov model says the match has a hidden state on every frame, home ball, away ball or dead, and that what we observe is a noisy consequence of it. Two tables define the model: transition probabilities between states, and an emission distribution for each state.

Three algorithms come with it, all exact, all in Rabiner’s tutorial. Forward-backward computes the probability of each state on each frame given the entire sequence. Viterbi finds the single most probable path. Baum-Welch fits the tables from unlabelled sequences by expectation maximisation.

hmmlearn implements all three behind a scikit-learn interface, with Gaussian, mixture and categorical emissions. The project describes itself as being in limited-maintenance mode: stable, not evolving.

Where it wins

  • Posteriors, not only labels. Forward-backward says how sure the model is on each frame, which a best path alone hides.
  • It can be fitted without labels, or run with tables set by hand from football knowledge.
  • Exact and cheap: linear in frames, quadratic in a state count of three.
  • A common language. Every sequence method in this family is an HMM with something relaxed.

What it costs

  • Emissions are assumed independent given the state. Football cues overlap heavily, and a naive emission model counts them twice.
  • Unsupervised fitting finds the states that explain the data best. They need not be H, A and D.
  • Spell lengths follow a geometric distribution, as in any first-order chain.
  • One emission family per state fits possession evidence poorly. Real evidence is a mix of distances, flags and absences.

How it plugs into possession

Mostly as mathematics. The decode at the heart of a possession pipeline is HMM inference whether or not this library is imported: states, switching costs, per-frame evidence, Viterbi. For three states the algorithm is a few lines of array code, and owning those lines keeps a build small and deterministic.

The part worth borrowing even then is forward-backward. Posterior confidence per frame separates the stretches a decoder knows from the stretches it bridged.

Our read

The reference model for sequence inference. Learn it from the library. For three states, write the decode yourself.

Sources

  1. hmmlearn: hidden Markov models in Python (repository)github.com
  2. A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition (Rabiner, 1989)doi.org