Probabilistic Programs as an Action Description Language

Actions description languages (ADLs), such as STRIPS, PDDL, and RDDL specify the input format for planning algorithms. Unfortunately, their syntax is familiar to experts only, not potential users of technology. Moreover, this limits ability describe complex large domains. We argue that programming (PLs), more specifically, probabilistic (PPLs), provide a suitable alternative. PLs are all programmers, support data types rich libraries manipulation, have powerful constructs, loops, sub-routines, local variables with which complex, realistic models objectives can be simply naturally specified. PPLs, make it easy distributions, essential stochastic models. The natural objection proposal opaque too expressive, making reasoning about them difficult. However, PPLs also come efficient inference algorithms, which, coupled growing body work on sampling-based gradient-based planning, imply execution monitoring carried out efficiently in practice. In paper, we expand proposal, illustrating its examples.