Particle filter

Particle filter class

class pyvib.ParticleFilter.ParticleFilter(N, R, Q, model, nStates, nParameters, measuredStates, resampleAlways=False, resampleDebug=False)

A particle filter class

Parameters

  • N (int) – Number of particles

  • R (float or array_like) – Variance of measured states len(R) == len(measuredStates)

  • Q (float or array_like) – Variance of actuation error Part of model

  • model (function(u, states, parameters, Q)) –

    Model that generates next step of states using previous states, parameters and Q statesDerivative can be used as a placeholder for the derivative Example:

    def model(u, states, parameters, statesDerivative, Q):

    m = parameters[:, 0] k = parameters[:, 1] c = parameters[:, 2] dt = 1.0 statesDerivative[:, 0] = states[:, 1] statesDerivative[:, 1] = 1.0/m*(-k*states[:, 0] - c*states[:, 1] + (u + randn(states.shape[0])*np.sqrt(Q)) states[:, 0] += statesDerivative[:, 0]*dt states[:, 1] += statesDerivative[:, 1]*dt

  • nStates (int) – Number of states in the system

  • nParameters (int) – Number of parameters in the system

  • measuredStates (int or array_like) – Which state number are measured Could be a single number or multiple in a list. Observation (z) must have the same length.

Methods

createGaussianParticles(mean, var)

Create gaussian distributed particles

createUniformParticles(ranges)

Create uniformly distributed particles

estimate()

Estimates true value and variance of states and parameters Results are saved in ParticleFilter.meanList and -.varList

getMeanAndVariance()

Get meanlist and varlist Mean and var

getPercentile(per)

Get the percentile of values

get_parameters()

Return the parameters of particles

get_states()

Return the states of particles

plotHistogram(column)

Plot histogram of a state or parameter

predict(u)

Predict state of next time step using control input

resample([thrScale])

Resamples particles IF necessary

simulateTrend(iterations, u[, percs])

Simulate the trend moving forward using the particles parameters and state

update(z[, debug])

Update the weights based on measurements and observation noise
createGaussianParticles(mean, var)

Create gaussian distributed particles

Parameters

  • mean (array_like) – Mean value of gaussian distributed guess len(mean) = nStates + nParameters

  • std (array_like) – Variation of gaussian distributed guess len(var) = nStates + nParameters

createUniformParticles(ranges)

Create uniformly distributed particles

Parameters

ranges (2D numpy.ndarray) – The uniform range of starting guess Shaped as [nStates + nParamteres, 2]

estimate()

Estimates true value and variance of states and parameters Results are saved in ParticleFilter.meanList and -.varList

getMeanAndVariance()

Get meanlist and varlist Mean and var

Returns

  • meanList (list of float 1D array) – Mean of each state for each time step

  • varList (list of float 1D array) – Variance of each state for each time step

getPercentile(per)

Get the percentile of values

Parameters

per (float) – Percentile <0.0, 1.0>

Returns

percentile – Percentiles

Return type

float 1D array

get_parameters()

Return the parameters of particles

Returns

parameters – Parameters of particles

Return type

float 2D array

get_states()

Return the states of particles

Returns

states – States of particles

Return type

float 2D array

plotHistogram(column)

Plot histogram of a state or parameter

Parameters

column (int) – Which column of self.particles should be pltted

predict(u)

Predict state of next time step using control input

Parameters

u (float or array_like) – The control input. Must follow rules of model function

resample(thrScale=0.5)

Resamples particles IF necessary

Parameters

thrScale (float, optional) – Thresholds for resampling scaled by number of particles

simulateTrend(iterations, u, percs=[0.5])

Simulate the trend moving forward using the particles parameters and state

Parameters

  • iterations (int) – Number of iterations to simulate

  • u (Data type defined by user in self.model) – Control input

  • percs (list of floats, optional) – Which percentile of parameters and states to simulate as true values

Returns

output – Simulation results output[i, j, k] gives the following i - the iteration j - the percentile k - the state number

Return type

float 3D array

update(z, debug=False)

Update the weights based on measurements and observation noise

zfloat or array_like:

The observation len(z) == len(measuredStates)