A convenient category of tracing measure kernels
Many probabilistic programming languages (PPLs) allow users to define custom proposals as programs. This requires conditionally evaluating a sampled proposal trace in the target program. Unfortunately, formulations of probability theory based on Markov categories as yet have no model of traces. This abstract introduces tracing and traced importance sampling in the higher-order Markov category QBS.
|A convenient category of tracing measure kernels (lafi23-final31.pdf)||629KiB|