Concept

The Probative Rapid Interactive Modeling Environment (PRIME) is a decision-support web application that provides modeling and reasoning capabilities intended to stretch the thinking of analysts and decision makers by producing a forecast of the plausible effects that could result from taking actions in a given situation. The plausibility of each forecast effect is explained by one or more structured arguments. The benefits are intended to be pedagogical (stretching a user's thinking via explanations of the identified plausible effects) rather than primarily prognostic (via predictive accuracy).

Modeling Capabilities

PRIME identifies plausible effects that could result from performing candidate actions at a site. Actions, for the purpose of PRIME modeling, are diplomatic, informational, military, or economic (DIME). A site includes one or more elements, each an entity or activity. An entity might be an actor (e.g., a person or organization), a physical entity (e.g., geographical, such as a region, country or village, or infrastructural, such as a building or road), or a conceptual entity (e.g., a religious or cultural icon). An activity might be occupational (e.g., fishing, trading) or social (e.g., communicating, voting). Each site element has one or more types (e.g., a bridge that is also a cultural icon) and zero or more relations. A relation in PRIME captures a connection between two site elements. For example, the relation has-leader could be used to connect a site element denoting a government (an organization) with another element denoting the current prime minister (a person). A site can also include descriptive profiles of its elements (e.g., to capture that a person is relatively affluent and has a post-graduate education).

To generate a forecast, PRIME uses a library of generic effects models (or rules) that describe plausible effects of actions on site elements. These include both the direct effects that actions can have on site elements as well as the indirect effects that can occur when changes in an element's state impact the state of related elements. In addition to the effects models, generating a forecast requires a model of the site where actions are going to be taken: the site elements and their relations and descriptive profiles comprise the model of the site. PRIME must also be given the candidate plan whose effects are to be forecast; this is simply a set of actions to be taken against specific site elements. A subject matter expert (SME) creates these objects in PRIME:

The models in PRIME rely on three foundational constructs as building blocks:

  1. Effects templates containing the dimensions or categories of interest for describing the effects on site elements
  2. Profile templates containing the attributes of interest for describing site elements (e.g., cultural variables for profiling socio-cultural entities)
  3. Taxonomies defining action types, relation types, and site-element types

Since the templates and taxonomies must reflect the essential characteristics of the area of interest, we anticipate that they will be edited to include the types of entities, activities, relations, actions, effects, and cultural variables of local interest. Given this ontological, procedural, and situational knowledge, PRIME generates forecasts, first based on direct effects, then moving to first-order indirect effects, second-order indirect effects, and so on. The user may review the forecast at any stage, inspecting explanations of forecast effects and optionally editing the forecast. An edited forecast can result in different models being matched in the successive rounds of indirect-effects generation.

Templates

When PRIME is initially deployed for a given area of interest, the provided templates for effects and profiling may need to be edited to better-fit local needs. Because the direct-effects models, indirect-effects models, and forecasts rely on these templates, the core distinctions included in the effects and profiling templates should be determined before creating site-specific direct or indirect effects models or forecasts.

The effects templates contain the categories of effects that are of interest, and for each category, they contain a set of subcategories (or variables) that will be modeled and forecasted. The existing effects template contains six categories: political, military, economic, social, infrastructure, and information (PMESII). Within the political category, for example, this template includes six subcategories of interest, leadership strength, political stability, secular influence, religious influence, ethnic influence, and external influence. This template provides the structure for all PMESII constraints and state descriptions used throughout PRIME. These are used in direct effects models (to describe plausible effects identified by the models), in indirect effects models (to describe generic conditions that constrain when the models apply as well as the plausible effects identified by the models), and in forecasts (to describe effects identified for specific site elements plausibly resulting from a course of action).

The profiling templates contain the attributes (i.e., non-relational variables) that are of interest for describing and reasoning about the site elements. For example, a cultural-profiling template captures a theoretical framework for describing socio-cultural entities. Each cultural variable can be thought of as a slider that is positioned between two extremes to describe a given socio-cultural entity (e.g., Hofstede's power distance index captures the extent to which subordinate members of a group accept and expect that power is distributed unequally). The profiling templates are used both to describe specific site elements and to capture generic conditions on effects models that constrain when they apply.

Taxonomies

PRIME is delivered with initial taxonomies for action types, relation types, and site-element types. Upon deployment, these taxonomies can be extended to better capture the local distinctions of interest. It is likely that taxonomy extension will be an ongoing modeling process that occurs in tandem with creation of effects models, sites, and plans.

Generic Models for Identifying Direct and Indirect Effects

As an on-going modeling activity PRIME is used to develop a library of rules for identifying both direct and indirect plausible effects. These rules are generic models in the sense that they describe effects on site-element types rather than on specific site elements. Consider a direct-effects model identifying plausible effects that an economic embargo would have on a country. The direct-effects model does not mention any specific country; rather, it identifies effects deemed plausible for any country on which an economic embargo is enforced. Hence, the model is termed generic. The generic models draw on the effects and profiling templates as well as the taxonomies. The generic models are intended to capture knowledge about how the world works, which may be commonsense knowledge or expert knowledge about effects propagation.

Specific Models for Capturing Sites and Plans

Two other important modeling tasks include: capturing site models and candidate plans. Each site model contains the specific entities and activities relevant to the courses of action being considered for a site of interest. They are captured using taxonomic types, relations (to connect pairs of site elements), and profiles (to describe the attributes of interest for each element). Each plan contains the actions to be taken on specific elements of a site model during a candidate course of action.

Reasoning Capabilities

PRIME produces a forecast of the plausible effects for a plan. This involves computing plausible values of the category and subcategory variables contained in the effects templates for elements in the site model. Values for the parent categories (e.g., political effects) are computed from the values of their subcategories (e.g., effects on leadership, effects on political strength, etc.). The values of the subcategories are computed for a site element from the plausible effects identified for that element by the applicable generic effects models.

Computing Direct Effects

PRIME generates a forecast for a plan by first considering its library of direct-effects models. These models describe the plausible effects on a site element from applying an action to that element. PRIME considers each site element that is the target of an action in the plan, and searches its library of direct-effects models to find all models relevant, given the element's type(s) and the model's action type(s). Next, the profile-template constraints are checked for each relevant model to determine which models are applicable. PRIME then merges the plausible effects from the applicable direct-effects models into an aggregated effects forecast for the site element (distinct aggregations are performed for each effects template).

The forecast effect is expressed as directions of change (e.g., increasing or decreasing) and (qualitative) magnitudes of change (e.g., little, moderate, or significant) to be expected for the subcategories contained in the effects templates. Each forecast effect includes an explanation for the expected value. For example, consider the forecast PMESII effects for providing foreign economic aid to a country: the political stability (effect subcategory) for that country is expected to increase moderately (plausible effect) because effective foreign aid, designed to improve overall conditions, will blunt opposition challenges and bolster confidence in the government (explanation). By merging the effects from all the applicable direct-effects models, PRIME creates a forecast that identifies plausible direct effects of the candidate plan for those site elements directly referenced in the plan.

Computing Indirect Effects

After the direct-effects forecast is completed, PRIME finds indirect effects that may propagate from the direct effects. For example, if a populace becomes discontented, the discontentment can reduce the political strength of the populace's government leader. PRIME considers all indirect-effects models in its library to find models relevant to the site and associated direct-effects forecast. PRIME then checks the conditions (e.g., the relations and profile constraints) of the relevant models to determine which are applicable. When PRIME finds an applicable model, the model's plausible effects are merged into the existing forecasts of the appropriate site elements. The form of the resulting forecast is identical to the direct-effects forecast. All effects added by applicable models, for both direct effects and indirect effects, are merged.

Since the relevance of indirect-effects models depends on the state of the forecast (in addition to the static associated types, relations, and descriptive profiles), indirect-effects forecasting proceeds in rounds. In each round, PRIME first finds all applicable indirect-effects models (given the current state of the forecast) and then merges the effects from the applicable models into the forecast. Note that the merging of effects changes the state of the forecast. The changed forecast could result in other indirect effects models becoming applicable. We term the effects found during the first round of matching as first-order indirect effects. The user may decide to request PRIME to do another round of indirect-effects forecasting. Any resulting effects from the second round are called second-order indirect effects. Likewise, the user may continue with additional rounds of forecasting, perhaps editing the results after each round. The results of successive rounds are likely less plausible than earlier rounds. A user of PRIME, who has special expertise in planning and examining courses of action, decides how many rounds of forecasting make sense. PRIME is intended to present plausible outcomes to the users to stretch their thinking. The hope is that PRIME will find and explain unexpected consequences, propagated effects that may have been overlooked otherwise. Users may edit plans and generate alternate forecasts after seeing these propagated effects.

Collaboration: User Feedback and External Forecasts

At any point during PRIME's forecasting cycle, others (users or software tools) can modify the emerging forecast. For example, a user might remove forecasted effects, modify effects, or add effects, and in so doing inject information that was not captured in PRIME's knowledge base. This could represent effect-propagation chains that have yet to be modeled, special circumstances that might mitigate modeled effects, or what ifs regarding effects that might be induced by outside agents. This forecast-editing capability flows directly from the collaborative nature of the underlying structured argumentation framework on which PRIME is built.

This same capability can be exploited to inject forecasted effects generated by other modeling and reasoning engines. When an external reasoner has newly forecasted effects, and they are communicated to PRIME using a shared effects template, they can be directly incorporated, merged with existing forecasts, and used as the basis for higher order effects forecasting as PRIME continues through its forecasting cycle.