Inference methods
TreePPL supports several inference methods.
The user needs to select the inference method and options.
To get a complete list of the available methods and options the user can type tpplc --help on the command line, which will provide all supported command-line options.
One can define the schema by issuing a command-line compilation command, or by making the appropriate choices in Python or R.
Please find below a reference list of the supported strategies with an explanation.
Importance Sampling
tpplc <program> -m is-lw
This is a basic inference schema that should not be used for complex models. It is useful for illustrating and understanding probabilistic programming.
Options available
--particles INT The number of particles (i.e., samples or iterations). The default is 5000
--no-early-stop Disables early stopping in certain inference algorithms.
--cps ARG Configuration of CPS transformation (only applicable to certain inference algorithms). The supported options are: none, partial, and full. Default: full.
--dynamic-delay Runs dynamic delayed sampling on the model.
--prune The model is pruned if possible.
Sequential Monte Carlo (SMC) Methods
TreePPL supports two SMC methods:
- bootstrap particle filter (BPF), and
- alive particle filter (APF).
Bootstrap particle filter (BPF)
tpplc <program> -m smc-bpf
This is the classical SMC schema. Except in special cases, the alive particle filter (APF) should be preferred as the BPF can suffer from path degeneration.
This inference is affected by the resample placement method.
Options available
--particles INT The number of particles (i.e., samples or iterations). The default is 5000
--subsample Whether to subsample the posterior distribution.
--subsample-size INT The number of subsamples to draw if --subsample is selected. Default: 1
--resample ARG The selected resample placement method, for inference algorithms where applicable. The supported methods are: likelihood (resample immediately after all likelihood updates), align (resample after aligned likelihood updates, forces --align), and manual (sample only at manually defined resampling locations). Default: manual
--resample-frac FLOAT Floating point number to trigger resampling for SMC-BPF when ESS is less than resampleFrac × particleCount. Default: 0.7
--cps ARG Configuration of CPS transformation (only applicable to certain inference algorithms). The supported options are: none, partial, and full. Default: full.
--prune The model is pruned if possible.
--dynamic-delay Runs dynamic delayed sampling on the model.
Alive particle filter (APF)
tpplc <program> -m smc-apf
This is a more advanced SMC schema than BPF. It avoids path degeneration by keeping a population of particles alive.
Reference: Kudlicka et al. 2020.
Options available
--particles INT The number of particles (i.e., samples or iterations). The default is 5000
--subsample Whether to subsample the posterior distribution.
--subsample-size INT The number of subsamples to draw if --subsample is selected. Default: 1
--resample ARG The selected resample placement method, for inference algorithms where applicable. The supported methods are: likelihood (resample immediately after all likelihood updates), align (resample after aligned likelihood updates, forces --align), and manual (sample only at manually defined resampling locations). Default: manual
--cps ARG Configuration of CPS transformation (only applicable to certain inference algorithms). The supported options are: none, partial, and full. Default: full.
--prune The model is pruned if possible.
Markov-chain Monte Carlo (MCMC) Methods
TreePPL supports a range of MCMC methods:
- Lightweight MCMC
- Naive and Trace MCMC
- particle MCMC--particle-independent Metropolis-Hastings
Lightweight MCMC
tpplc <program> -m mcmc-lightweight
This is the Markov-chain Monte Carlo (MCMC) schema that should be used in most cases.
Moves are proposed by resampling one assume statement at random.
Draws that are aligned together with some unaligned draws are reused to save computational effort.
It has an option of whether to do a continuation-passing style transformation in order to avoid recomputation of the likelihood. For details, see Lunden et al. 2024.
Options available
--align Whether or not to align the model for certain inference algorithms.
--debug-iterations Output various debug information each iteration.
--sampling-period N Sample the mcmc-chain every Nth iteration. Default: 1
--incremental-printing Print each sample as it is produced instead of at the end.
--particles INT The number of particles (i.e., samples or iterations). The default is 5000
--mcmc-lw-gprob FLOAT The probability of performing a global MH step (non-global means only modify a single sample in the previous trace). Default: 0.1
--kernel Use drift Kernel in MCMC.
--drift FLOAT Floating point number which corresponds to the standard deviation (sigma) of the normal distribution that will be used for the automatic drift kernel. Default: 1.
--cps ARG Configuration of CPS transformation (only applicable to certain inference algorithms). The supported options are: none, partial, and full. Default: full.
--debug-alignment-html FILE Output an interactive .html file showing alignment results to the given file.
Naive and Trace MCMC
tpplc <program> -m mcmc-naive
tpplc <program> -m mcmc-trace
The new proposed sample is a complete rerun of the entire simulation. Those are pedagogical methods that should not be used for real problems.
Options available
--particles INT The number of particles (i.e., samples or iterations). The default is 5000
Particle MCMC--Particle independent Metropolis-Hastings (PMCMC-PIMH)
tpplc <program> -m pmcmc-pimh
This is perhaps the simplest version of an SMC-within-MCMC approach, where each proposal in the MCMC is a full set of samples (particles) generated by SMC.
Options available
--pmcmcParticles INT The number of particles for the smc proposal computation. The default is 2
--particles INT The number of particles (i.e., samples or iterations). The default is 5000
--cps ARG Configuration of CPS transformation (only applicable to certain inference algorithms). The supported options are: none, partial, and full. Default: full.