A framework for development, teaching and deployment of inference algorithms

We present symfer, a software framework for probabilistic inference algorithms. Each inference algorithm (like variable elimination, junction tree propagation, recursive conditioning) is represented as a symbolic manipulation of factor algebra expressions. In combination with the readability and terseness of Python code, this uniform representation makes the framework very suitable for teaching, as well as for explorative research: using the interactive Python interpreter, one can combine features of different algorithms and examine their effect. Numeric evaluation happens in a separate stage, implemented in Java/C for efficient execution and high portability. We exploit the latter feature in an application that performs inference on a smartphone.