Probabilistic Effekt

Inference programming - the adaptation of inference algorithms to a particular model or problem - is difficult in languages that do not deal with side effects or do not handle side effects properly due to their nature of relying on effectful computation. In functional programming, monad transformers are the most common approach to control side effects. However, this makes it difficult to customise such algorithms. Effects and effect handlers on the other hand introduce another possibility to deal with programmable inference. This thesis will demonstrates this idea by developing three inference patterns (Metropolis-Hastings, slice sampling and rejection sampling) using the language Effekt. The result is a more user-friendly way to adapt and combine parts of these algorithms


Jonathan Brachthäuser

Philipp Schuster