Home / Curated / Model Families / Hidden Markov decision tree

Hidden Markov decision tree

Model Families AIMech:HiddenMarkovDecisionTree DRAFT

Aliases: HMDT

A conditional latent sequence-model family that extends a probabilistic hierarchical decision tree by coupling each decision level to the same level at the previous time.

Identity Notes & Concept Scope The Jordan–Ghahramani–Saul model uses probabilistic choices down a hierarchy and Markov persistence within each level. It is not a hard deterministic classification tree, an independent-chain factorial HMM, a tree-shaped index, or a particular variational inference algorithm. Current higher-level decisions influence lower-level decisions; independent factorial chains arise only under additional restrictions. The model conditions on an external input sequence. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 379 preceding native IDs/names/aliases for the four researched candidate names. Both full-checkout fixed-string searches included ignored and hidden files, with 7 and 7 matching lines. All 6 unique matched contents were read: pending leads and preceding source/review context, not existing records of these three new identities. Expanded source IDs, plurals, Gibbs-machine and mixture-transition terminology were included. All 63 semantic label/synonym hits, 17 exact fully annotated AIO CLASS nodes with outgoing edges, their generated YAMLs and six native neighbors were read or byte-verified after prior complete reading. Markov mixtures of experts substantially overlaps the existing IOHMM family, so no fourth ID or unconditional alias is created. Absence is bounded to the pinned vocabulary and inspected neighborhood.

Claims & Grounded Evidence

1 Claim

Mechanism #method

For level m and time t, its latent decision z_t^m is distributed conditional on the current input x_t, all current ancestor decisions z_t^1,...,z_t^(m-1), and its own preceding decision z_(t-1)^m. Initial decisions condition on the input and current ancestors without a preceding-time state. Multiplying these normalized conditional factors and one observation factor p(y_t|x_t,z_t^1,...,z_t^M) per time gives a normalized directed model conditional on inputs. A latent path through the hierarchy selects an output expert. Exact inference may enumerate the joint states but grows exponentially with hierarchy depth. The paper explores three distinct approximations: a product of chain-path distributions that preserves temporal dependence within each level; a product over times of tree-structured distributions that preserves current ancestor dependence; and a point-mass path approximation. Its point-mass procedure optimizes one level’s path by a Viterbi subroutine while holding the other paths fixed. These are inference choices for the family, not alternative definitions of its identity.

supports Source: Hidden Markov Decision Trees · Locator: Sections 2.1–2.3 and Figure 3, printed pp.502–504 / PDF pp.2–4; Sections 3.1–3.4 and Figures 4–5, pp.504–506 / PDF pp.4–6; Sections 4–5 and Table 1, pp.506–507 / PDF pp.6–7

⚠️ Curated Limitations & Negative Results

Same-level temporal connections and current-ancestor dependencies are defining restrictions of the inspected construction. They do not cover arbitrary cross-time connections or all models whose names contain hidden Markov tree.
The probabilistic hierarchy sums over decision paths; it is not necessarily a hard classification tree. Input conditioning and an output expert do not imply a causal interpretation of the graph.
Independent latent chains, a single HMM and an ordinary hierarchical mixture arise under specific restrictions, not as unconditional aliases of this family.
Forest-of-chains, forest-of-trees and point-mass inference discard different posterior dependencies. A coordinate Viterbi step finds the best path for one fixed-context chain, not a guaranteed global joint MAP path.
The paper motivates hierarchical time scales but its synthetic runs did not always recover the intended hierarchy. No universal coarse-to-fine organization, general convergence guarantee, benchmark advantage or author-code reproduction is asserted. Detailed generic variational updates remain a separate source lead.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:BayesianNetwork` Bayesian Network	broad_match	The finite time-unrolled conditional factorization is a directed acyclic probabilistic network with inputs conditioned on. The pinned BayesianNetwork target is broader than this hierarchical temporal model.
`AIO:DecisionTree` Decision Tree	related_match	The model extends probabilistic tree decisions with time dependence. The pinned target is defined and parented as a classification task; a conditional sequence model need not be a classifier, so this is a structural association rather than a model-to-task subclass.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`method-source`	Hidden Markov Decision Trees Michael I. Jordan, Zoubin Ghahramani, Lawrence K. Saul	primary paper full_text	ID: https://proceedings.neurips.cc/paper/1996/hash/6c8dba7d0df1c4a79dd07646be9a26c8-Abstract.html sha256:e073f1c78b48…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T20:01:01Z

Updated: 2026-10-11T20:01:50Z

Source File: data/curated/HiddenMarkovDecisionTree.yaml