k-order Dynamic Bayesian Networks (k-TBN)

A dynamic Bayesian network models a stochastic process by describing how the distribution over a set of variables at time \(t\) depends on the past. The classical 2-TBN limits that dependency to a single step backward (\(t-1\)). A k-order dynamic Bayesian network (k-TBN) generalizes this: a variable at time \(t\) may depend on any of the \(k\) most recent time slices \(t, t-1, \ldots, t-k+1\).

Rather than storing an (infinite) unrolled network, a k-TBN stores a compact template made of exactly \(k\) time slices. This template captures both the initial distribution (slices \(0, \ldots, k-2\)) and the transition kernel (slice \(k-1\), reused for every \(t \geq k-1\) since the process is time-homogeneous).

Two kinds of variables are distinguished:

temporal variables (processes), which evolve through time and are therefore represented by \(k\) instances (one per slice) in the template;
atemporal variables, which are constant through time (e.g. a static context parameter) and are represented by a single instance.

Internally, a temporal node is named with bracket notation: the \(t\)-th instance of a process base is base[t] (e.g. \"X[0]\", \"X[1]\"). Atemporal variables keep their bare name. Most of the public API lets you address a node either that way, or through an explicit (base, slice) pair, slice being pyagrum.ktbn.KTBN.ATEMPORAL (-1) for an atemporal variable.

A minimal example

import pyagrum as gum
import pyagrum.ktbn as ktbn

model = ktbn.KTBN(2)                      # order k=2
model.addTemporal("X[2]")                 # a binary process
model.addAtemporal("C[2]")                # a binary static context
model.addArc("X", 0, "X", 1)              # X[0] -> X[1]
model.addArc("C", ktbn.KTBN.ATEMPORAL, "X", 0)
model.generateCPTs()                      # or fillCPT(...) node by node

bn10 = model.unroll(10)                   # a plain BayesNet over 10 slices

ie = ktbn.KTBNInference(model)
ie.addObservation("X", 1, 1)              # observe X[1] = 1
ie.addIntervention("C", ktbn.KTBN.ATEMPORAL, 0)   # do(C = 0)
ie.addTarget("X")
ie.makeInference(5)
print(ie.posteriors("X"))                 # P(X[t] | ...) for t = 0..4

Tutorial

Tutorial on k-order dynamic Bayesian networks

Input / Output

k-TBNs can be saved using the native JGUM / BGUM Format Reference.

ktbn.saveKTBN(model, "model.jgum")   # jgum (JSON)
ktbn.saveKTBN(model, "model.bgum")   # bgum (binary)

Reference