lppl

(Disclaimer – this page is written for geeks like me. A more publicly accessible explanation of the lppl vision is coming soon.)

Open-universe probabilistic programming in modern C++, done in exactly the idiosyncratic way that I want to do it.

• Record-based (which essentially means trace-based, but we don’t hide the tracing from you)
• Sample-based, with posterior queries by default computed online rather than after sampling is complete – defaulting to constant memory complexity
• Explicit PRNG state (reproducibility)
• Algebraic sum types everywhere instead of boxing

This should look familiar.

template<size_t N>
std::tuple<double, double> 
linear_regression(record_t<Normal, Gamma>& r, std::shared_ptr<data_1d<N>> data) {
    auto intercept = sample(r, "intercept", Normal(), rng);
    auto slope = sample(r, "slope", Normal(), rng);
    auto scale = sample(r, "scale", Gamma(), rng);
    for (size_t ix = 0; ix != N; ix++) {
        observe(
            r,
            "obs/" + std::to_string(ix),
            Normal(data->x[ix] * slope + intercept, scale),
            data->y[ix]
        );
    }
    return std::make_tuple(intercept, slope);
}

lppl is open-source and available at this webpage. There’s lots of documentation! You can also download the latest tagged version of the source, or see all versions.

⚠️ Versions of lppl greater than v0.9.0 have changed licenses from LGPL3 to GPL3. Please contact us if you are interested in a license exception.

Details

• About
• Examples
• Tests

Associated tools

• glppl transforms lppl programs into probability distributions over directed graphical models.

Inference

lppl supports the following inference algorithms:

• importance sampling
  • likelihood weighting
  • generic importance sampling with user-defined proposal
• metropolis-hastings
  • ancestor metropolis
  • generic metropolis with user-defined proposal
• generic filtering using any inference algorithm to compute updates

More algorithms to come…

Coming attractions

• optionally zero dynamic allocation (yes, even in open-universe models)
• other inference algorithms (CSIS, ADVI, …)
  • some inference algorithms may end up going in separate libraries, since they might depend on external libraries (e.g., Adept for neural network-based inference). This is all very hypothetical anyway.
• more query types, and composability of queries

The lppl webpage and its subpages are licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.