Imprecise hiddenMarkovmodels(iHMMs)areanextensionof
standardHMMsthatarguablyallowforproperhandlingofthe
imprecisionintheparametersthatariseinmanydomains.Inthis
paper,wepresentedalgorithmsforstandardusagesuchascom-
puting likelihood,performing filtering/prediction,and finding
optimisticandpessimisticmostlikelystatesequences.Whenthe
parametersarespecified asinterval-valuedprobabilities,allalgo-
rithms runinquadratictimeinthenumberofstatesandlinearin
the numberoftimesteps.Remarkably,thisisthesametime
complexityoftheanalogousalgorithmsforstandardHMMs.When
imprecisiontakesamorecomplexform(e.g.assetsoflinear
inequalities),thetimecomplexitygrowsonlybythetimeofsol-
ving alinearprogramofsizelinearintheinput.
Experiments withrealdatashowedthatiHMMscanbeusedas
“cautious” classifiers thatsuspenddecisionmakingwhenthereisnot enoughstatisticalevidencetoconfidently supportadecision.
In addition,iHMMscanserveasvaluabletoolstoperformanalysis
of thesensitivityofprecise(non-stationary)HMMstovariationsof
the parameters.
The imprecisioninthenumericalparametersofthemodel
translatestoindeterminacywhenusingthemodelstomakedeci-
sions asintheapplicationsweshow.Wehaveadoptedhereinterval
dominance asthebasecriteriontosuspendjudgment.Theliterature
containsothercriteriathatareworthevaluating,suchasmaximality
andE-admissibility.Implementingsuchcriteriawillrequiredevel-
oping efficientalgorithms.Weleavethatasfuturework.