PRIME's development draws upon concepts and software from two other key systems previously developed by SRI: the Structured Evidential Argumentation System (SEAS) and the The Link Analysis Workbench (LAW). SEAS provides the foundation for employing structured arguments to capture generic effects models and for aggregating forecasted effects based on those models or from other reasoning engines. LAW provides the foundation for matching the requisite conditions of indirect-effects models to an existing forecast, to determine which models apply and have their effects added to the forecast.
SRI has been investigating the use of template-based structured argumentation as a means of capturing and guiding collaborative analysis. The idea is to capture best analytic practices for a given class of problems in a template and then use that template as the basis for collecting evidence and drawing conclusions about specific situations. Unlike other work focused on automating human uncertain reasoning, this approach focuses on recording and coordinating human reasoning. A key aspect of this has been the use of graphical depictions of arguments to rapidly convey the state of lines of inquiry, from evidence to conclusion, highlighting information needs as well as the evidence that drives the conclusion. To support this approach, SRI created a collaborative software tool called SEAS (Structured Evidential Argumentation System). Using this tool, contributing analysts directly manipulate depictions of arguments, adding and interpreting evidence relative to questions raised by the template, debate and draw conclusions based on the collective evidence, and finally use these depictions to convey their findings to decision makers.
PRIME's indirect-effects inference capability was developed based upon techniques drawn from LAW (Link Analysis Workbench. LAW is an analysis tool designed to capture and match patterns of interest, in large sets of relational data. The patterns are represented as semantically labeled networks of connected entities, where the connections represent specific types of relationships among specific types of entities, making LAW patterns ideally suited for modeling the propagation of indirect effects among related entities (e.g., if public discontentment within a given group increases, then the strength of that group's leadership will suffer). A pattern in LAW consists of a number of typed entities, with typed relationships among them, and with constrained attribute values on the entities. A match for a given pattern is a subgraph within a set of relational data that includes specific entities/relationships with matching types and whose attribute values fall within the specified constraints. We use such patterns in PRIME to describe the conditions under which (e.g., PMESII) effects will propagate across related entities. Such patterns constitute the applicability conditions of generic indirect-effects models.
At the heart of the LAW system is a graph-based pattern representation and matching capability. This includes a flexible, hierarchical pattern representation language based on graphs, a pattern comparison metric based on a variant of graph edit distance, and an anytime search mechanism for finding approximate matches to the pattern in large datasets. LAW's current matching algorithm for finding patterns in the data is based on A* search. The search process is designed to find a good set of pattern matches quickly, and then use those existing matches to prune the remainder of the search.
By building PRIME on top of this SRI infrastructure, previously established to support SEAS and LAW, PRIME inherits, or is poised to inherit, many advanced capabilities, including collaborative modeling, full access control, information assurance, and taxonomy management.