On the Applicability of Probabilistic Programming Languages for Causal Activity Recognition

Recognizing causal activities of human protagonists, and jointly inferring context information like location of objects and agents from noisy sensor data is a challenging task. Causal models can be used, which describe the activity structure symbolically, e.g. by precondition-effect actions. Recently, probabilistic programming languages (PPLs) arose as an abstraction mechanism that allow to concisely define probabilistic models by a general-purpose programming language, and provide off-the-shelf, general-purpose inference algorithms. In this paper, we empirically investigate whether PPLs provide a feasible alternative for implementing causal models for human activity recognition, by comparing the performance of three different PPLs (Anglican, WebPPL and Figaro) on a multi-agent scenario. We find that PPLs allow to concisely express causal models, but general-purpose inference algorithms that are typically implemented in PPLs are outperformed by an application-specific inference algorithm by orders of magnitude. Still, PPLs can be a valuable tool for developing probabilistic models, due to their expressiveness and simple applicability.