29 october 2017: A new version (0.7.1) has been released.
What is RankPL?
RankPL is a semi-qualitative probabilistic programming language aimed at modelling uncertainty expressible by distinguishing normal from surprising events. For example, we can express that a statement is only normally or only surprisingly executed, or that a variable normally becomes X but may surprisingly become Y.
This kind of uncertainty often appears in common sense reasoning problems. For example, if we diagnose a boolean circuit, we assume that each component normally functions correctly, or if we process sensor data, we assume that each piece of data we receive is normally correct. If precise probabilites of such events are unknown or irrelevant, a probabilstic approach is inappropriate.
Like a regular probabilistic programming language, RankPL can be used to easily represent models that generate explanations and predictions based on observed data. Executing a RankPL program produces a ranking over possible outcomes: first the normal outcomes, then the surprising ones, then the very surprising ones, and so on. The computational simplicity of ranks allows for efficient and exact inference without the need for sampling techniques.
Semantically, RankPL is based on ranking theory. This is a formalism for reasoning about uncertainty in which events receive discrete, integer-valued ranks that represent degrees of surprise. Ranking theory can be understood as a qualitative abstraction of probability theory that retains many desirable features of the probabilistic approach. For example, it supports update through conditioning and representation of complex (in)dependence relationships. Ranking theory is well known as a basis for logic-based AI approaches such as belief revision and non-monotonic reasoning.
A detailed introduction and language reference can be found in the wiki.
RankPL is described in the paper "RankPL: A Qualitative Probabilistic Programming Language" (presented at ECSQARU 2017). Note that the syntax described in this paper is outdated and that the latest version of RankPL contains many new features. Consult the wiki for an up-to-date description. An updated and extended version of the paper is in preparation.
The RankPL interpreter is distributed as a self-contained jar file
RankPL.jar. The jar file for the latest release can be downloaded here. Basic usage is:
java -jar RankPL.jar -source source_file.rpl
source_file.rpl is the program to run. Command line options are discussed in the wiki.
RankPL is built using Maven. The process is straightforward. For example, to build the self-contained Jar file, run the following command from the project's root directory:
The self-contained Jar file will be written to
Please contact me if you have questions or suggestions (email@example.com). Contributions are welcome.
RankPL is licensed under the MIT license.