Inference
TreePPL supports a continuously growing number of inference schemas.
To define a schema, the user has to select the inference method and options.
To get a complete list of the available methods and options the user can type tpplc 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 -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.
Sequential Monte Carlo (SMC) Methods​
TreePPL supports two SMC methods:
- bootstrap particle filter (BPF), and
- alive particle filter (APF).
Both inference methods have a resampling option.
Bootstrap particle filter (BPF)​
tpplc -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.
Alive particle filter (APF)​
tpplc -m smc-apf
This is the more advanced SMC schema. It avoids path degeneration by keeping a population of particles alive.
Reference: Kudlicka et al. 2020.
Likelihood resampling​
tpplc -m smc-bpf --resample likelihood
resamples the particles at every observe, weight, and logWeight statement.
Aligned resampling​
tpplc -m smc-bpf --resample align
resamples the particles after observe, weight, and logWeight statements that are aligned.
Manual resampling​
tpplc -m smc-bpf --resample manual
resamples only at the resample keyword. This is the default behavior.
Markov-chain Monte Carlo (MCMC) Methods​
TreePPL supports a range of MCMC methods:
- lightweight MCMC
- aligned lightweight MCMC (with different options for CPS)
- naive and trace MCMC
- particle MCMC--particle-independent Metropolis-Hastings
Lightweight MCMC​
This is not as efficient as its aligned version and should be used only for testing.
Aligned lightweight MCMC​
tpplc -m mcmc-lightweight --align
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:
tpplc -m mcmc-lightweight --cps partial
is the default, which should be the most performant. For details, see Lunden et al. 2024.
One can also use a full CPS transformation
tpplc -m mcmc-lightweight --cps full
which is the traditional algorithm, or to turn it off completely
tpplc -m mcmc-lightweight --cps none
which may also be beneficial in certain scenarios, where the CPS overhead is not a good trade-off vs recomputing the likelihood.
Naive and trace MCMC​
The new proposed sample is a complete rerun of the entire simulation. Those are pedagogical methods that should not be used for real problems.
Particle MCMC--Particle independent Metropolis-Hastings (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.