Robust, fast, fuss-free MCMC parameter estimation from .net, R, or MatLab
Note: This research project has reached its conclusion. These pages are maintained for reference and archival purposes.
Filzbach is a flexible, fast, robust, parameter estimation engine that allows you to parameterize arbitrary, non-linear models, of the kind that are necessary in biological sciences, against multiple, heterogeneous data sets. Filzbach allows for Bayesian parameter estimation, maximum likelihood analysis, priors, latents, hierarchies, error propagation and model selection, often with just a few lines of code.
Scientific results enabled by Filzbach have been published in tens of peer-reviewed publications. For more details see the list of publications below, and the Computational Science Lab tools page.
You can also build Filzbach library from sources (GitHub) (opens in new tab) using GNU g++ or Microsoft Visual C++.
Inspiration
Traditionally, ecology and biology has been largely split into the purely empirical (generating or analysing data with only informal use of models) and the purely theoretical (analysing ideas-rich models that have been, at best, only informally constrained with data). However, to create a precise, predictive, understanding of ecological and biological systems it is necessary to bridge this gap, using data to formally parameterize, and select between, aribtrary, ideas-rich models.
Features
- Specify parameters, define the likelihood, and Filzbach does the rest
- The automatic adaptive MCMC sampling algorithm copes with a wide range of different problems with no need for any manual tuning
- Automatic handling of multiple chains, testing for convergence, calculation of MLEs, posterior means and credible intervals on all parameters; and AIC, BIC, DIC
- Easy error propagation of parameter uncertainty through any model
- Fast and robust compared to commonly used alternatives
- Comes with a library of easy to use parameter distributions — but can be extended include any others, so long as they are written in C
- FilzbachR makes the power of Filzbach available through R, and crucially, allows the user to specify the model, and the likelihood, in R itself. This allows for Filzbach analyses that use any of the stats and libraries already available in R
- Similarly, FilzbachMatlab will make the power of Filzbach available in MatLab. FilzbachMatLab is not yet released, but watch this space, and feel free to contact us for more details
To try online, download, and learn more, see the Filzbach section of our new tool site.
Filzbach was developed by Drew Purves and Vassily Lyutsarev, within the Computational Science Lab at Microsoft Research, Cambridge.