internal static void Run()
{
// http://ocw.mit.edu/courses/aeronautics-and-astronautics/16-410-principles-of-autonomy-and-decision-making-fall-2010/lecture-notes/MIT16_410F10_lec21.pdf
var states = new Registry <IState>();
var x1 = states.Add(new NamedState("x1"));
var x2 = states.Add(new NamedState("x2"));
var x3 = states.Add(new NamedState("x3"));
var observations = new Registry <IObservation>();
var o2 = observations.Add(new NamedObservation("o2"));
var o3 = observations.Add(new NamedObservation("o3"));
// test the HMM with a known graph
{
var initial = new InitialStateMatrix(states);
initial.SetProbability(x1, 1);
initial.SetProbability(x2, 0);
initial.SetProbability(x3, 0);
var transitions = new TransitionMatrix(states);
transitions.SetTransition(x1, x1, 0);
transitions.SetTransition(x1, x2, 0.5);
transitions.SetTransition(x1, x3, 0.5);
transitions.SetTransition(x2, x1, 0);
transitions.SetTransition(x2, x2, 0.9);
transitions.SetTransition(x2, x3, 0.1);
transitions.SetTransition(x3, x1, 0);
transitions.SetTransition(x3, x2, 0);
transitions.SetTransition(x3, x3, 1);
var emissions = new EmissionMatrix(states, observations);
emissions.SetEmission(x1, o2, 0.5);
emissions.SetEmission(x2, o2, 0.9);
emissions.SetEmission(x3, o2, 0.1);
emissions.SetEmission(x1, o3, 0.5);
emissions.SetEmission(x2, o3, 0.1);
emissions.SetEmission(x3, o3, 0.9);
var hmm = new HiddenMarkovModel(states, initial, transitions, emissions);
// expected output: 1, 3, 3, 3, 3, 3, 3, 3, 3
hmm.ApplyViterbiAndPrint(new[] { o2, o3, o3, o2, o2, o2, o3, o2, o3 });
// expected output: 1, 2, 2, 2, 2, 2, 2, 2
hmm.ApplyViterbiAndPrint(new[] { o2, o3, o3, o2, o2, o2, o3, o2 });
}
}
private void ValidateParameters()
{
if (EmissionMatrix == null || TransitionMatrix == null || SequenceOfObservations == null ||
StateSpace == null || ObservationSpace == null)
{
throw new ArgumentException("Parameters cannot be null");
}
if (ObservationSpace.Length != EmissionMatrix.GetLength(1) || ObservationSpace.Length == 0)
{
throw new ArgumentException("N should be greater than 0 and consistent");
}
#if _USE_ARRAYS_INSTEAD_OF_MATRIX_HASHTABLE
if (StateSpace.Length != TransitionMatrix.GetLength(0) || TransitionMatrix.GetLength(0) != TransitionMatrix.GetLength(1) || TransitionMatrix.GetLength(1) != EmissionMatrix.GetLength(0) || EmissionMatrix.GetLength(0) != InitialProbabilitiesOfStates.Length || StateSpace.Length == 0)
{
throw new ArgumentException("K should be greater than 0 and consistent");
}
#else
if (StateSpace.Length != TransitionMatrix.GetLength(0) ||
TransitionMatrix.GetLength(0) != TransitionMatrix.GetLength(1) ||
TransitionMatrix.GetLength(1) != EmissionMatrix.GetLength(0) ||
EmissionMatrix.GetLength(0) != InitialProbabilitiesOfStates.Count || StateSpace.Length == 0)
{
throw new ArgumentException("K should be greater than 0 and consistent");
}
#endif
if (SequenceOfObservations.Length == 0)
{
throw new ArgumentException("T should be greater than 0 and consistent");
}
if (
StateSpace.Select(state => ObservationSpace.Sum(observation => EmissionMatrix[state, observation]))
.Any(sum => sum <= 0.99 || sum >= 1.01))
{
throw new ArgumentException("EmissionMatrix has not normalized probabilities"); //Exception("Emi");
}
if (
StateSpace.Select(state1 => StateSpace.Sum(state2 => TransitionMatrix[state1, state2]))
.Any(sum => sum <= 0.99 || sum >= 1.01))
{
throw new ArgumentException("TransitionMatrix has not normalized probabilities"); //Exception("Emi");
}
}
internal static void Run()
{
// http://www.cs.sfu.ca/~anoop/teaching/CMPT-413-Spring-2014/index.html
const double compareEpsilon = 0.000001D;
var states = new Registry <IState>();
var adjective = states.Add(new NamedState("A"));
var noun = states.Add(new NamedState("N"));
var observations = new Registry <IObservation>();
var clown = observations.Add(new NamedObservation("clown"));
var killer = observations.Add(new NamedObservation("killer"));
var crazy = observations.Add(new NamedObservation("crazy"));
var problem = observations.Add(new NamedObservation("problem"));
// test the HMM with a known graph
{
var initial = new InitialStateMatrix(states);
initial.SetProbability(adjective, 1 / 3D);
initial.SetProbability(noun, 2 / 3D);
var transitions = new TransitionMatrix(states);
transitions.SetTransition(adjective, adjective, 0);
transitions.SetTransition(adjective, noun, 1);
transitions.SetTransition(noun, adjective, 0.5);
transitions.SetTransition(noun, noun, 0.5);
var emissions = new EmissionMatrix(states, observations);
emissions.SetEmission(adjective, clown, 0);
emissions.SetEmission(adjective, killer, 0);
emissions.SetEmission(adjective, problem, 0);
emissions.SetEmission(adjective, crazy, 1);
emissions.SetEmission(noun, clown, 0.4);
emissions.SetEmission(noun, killer, 0.3);
emissions.SetEmission(noun, problem, 0.3);
emissions.SetEmission(noun, crazy, 0);
var hmm = new HiddenMarkovModel(states, initial, transitions, emissions);
// P(AA | killer clown)
var paa = hmm.GetProbability(killer.As(adjective), clown.As(adjective));
Debug.Assert(Math.Abs(paa) < compareEpsilon);
// P(AN | killer clown)
var pan = hmm.GetProbability(killer.As(adjective), clown.As(noun));
Debug.Assert(Math.Abs(pan) < compareEpsilon);
// P(NN | killer clown)
var pnn = hmm.GetProbability(killer.As(noun), clown.As(noun));
Debug.Assert(Math.Abs(0.04 - pnn) < compareEpsilon);
// P(NA | killer clown)
var pna = hmm.GetProbability(killer.As(noun), clown.As(adjective));
Debug.Assert(Math.Abs(pna) < compareEpsilon);
}
// test supervised learning of the HMM
{
var trainingSet = new List <IList <LabeledObservation> >
{
new[] { killer.As(noun), clown.As(noun) },
new[] { killer.As(noun), problem.As(noun) },
new[] { crazy.As(adjective), problem.As(noun) },
new[] { crazy.As(adjective), clown.As(noun) },
new[] { problem.As(noun), crazy.As(adjective), clown.As(noun) },
new[] { clown.As(noun), crazy.As(adjective), killer.As(noun) },
};
// prepare the matrices
var initial = new InitialStateMatrix(states);
var transitions = new TransitionMatrix(states);
var emissions = new EmissionMatrix(states, observations);
// learn the probabilities
var learner = new ClassicalBaumWelchLearning();
learner.Learn(initial, transitions, emissions, trainingSet);
var hmm = new HiddenMarkovModel(states, initial, transitions, emissions);
// P(AA | killer clown)
var paa = hmm.GetProbability(killer.As(adjective), clown.As(adjective));
Debug.Assert(Math.Abs(paa) < compareEpsilon);
// P(AN | killer clown)
var pan = hmm.GetProbability(killer.As(adjective), clown.As(noun));
Debug.Assert(Math.Abs(pan) < compareEpsilon);
// P(NN | killer clown)
var pnn = hmm.GetProbability(killer.As(noun), clown.As(noun));
Debug.Assert(Math.Abs(0.04 - pnn) < compareEpsilon);
// P(NA | killer clown)
var pna = hmm.GetProbability(killer.As(noun), clown.As(adjective));
Debug.Assert(Math.Abs(pna) < compareEpsilon);
// apply the viterbi algorithm to find the most likely sequence
hmm.ApplyViterbiAndPrint(new[] { killer, crazy, clown, problem });
hmm.ApplyViterbiAndPrint(new[] { crazy, killer, clown, problem });
hmm.ApplyViterbiAndPrint(new[] { crazy, clown, killer, crazy, problem });
var p = hmm.Evaluate(new[] { killer, crazy, clown, problem });
}
}
