Make HierarchicalSSM compatible with the BF #70
Conversation
There was a problem hiding this comment.
I worry that we add some bloat by storing a cloud of Rao-Blackwellised particles. Consider a user who designed a summary statistic f(x::ParticleDistribution). The internals of that function would need to be rewritten to support the special case of ParticleDistribution{RaoBlackwellisedParticle{XT,ZT},...}.
Regardless, this is a relatively specific concern for an elegant fix. I approve.
| function filter( | ||
| rng::AbstractRNG, | ||
| model::HierarchicalSSM, | ||
| alg::BootstrapFilter, | ||
| observations::AbstractVector; | ||
| ref_state::Union{Nothing,AbstractVector}=nothing, | ||
| kwargs..., | ||
| ) | ||
| ssm = StateSpaceModel( | ||
| HierarchicalDynamics(model.outer_dyn, model.inner_model.dyn), | ||
| HierarchicalObservations(model.inner_model.obs), | ||
| ) | ||
| return filter(rng, ssm, alg, observations; ref_state=ref_state, kwargs...) | ||
| end |
There was a problem hiding this comment.
I love Julia for this
|
That's a fair point, though note that we can't assume that the outer state is a vector (e.g. it could be the jump collection of a Lévy process) so we can't just concatenate them together as vectors. |
Solves #34.
A middle-ground approach to making the HierarchicalSSM compatible with the bootstrap filter. I've add two special dynamics types and when
filteris called withBootstrapFilterandHierarchicalSSM, the model is converted to a regularStateSpaceModelusing these two special types.I think a long-term ideal approach would be to introduce methods for
dynamics()andobservations()which construct the dynamics/observation process on the fly so that the inner functionsupdate,predictcan act on the HierarchicalSSM directly. I found this ended up being quite faffy though and require some definitions on what exactly anAbstractStateSpaceModelis which seems like a bit too big of a change to worry about right now (maybe later if we include mixed linear/non-linear models [1].What do you think @charlesknipp? Fine for now?
[1] F. Lindsten, P. Bunch, S. J. Godsill and T. B. Schön, "Rao-Blackwellized particle smoothers for mixed linear/nonlinear state-space models," 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 2013, pp. 6288-6292, doi: 10.1109/ICASSP.2013.6638875. keywords: {Trajectory;Smoothing methods;Monte Carlo methods;Approximation methods;Joints;Vectors;State-space methods;Rao-Blackwellization;particle smoothing;backward simulation;sequential Monte Carlo},