Using SIPE-2 to integrate planning for military air campaigns

by Thomas J. Lee and David E. Wilkins, SRI International

In 1995 and 1996, SRI participated in an integrated feasibility demonstration to show, in an operational environment, the relevance of generative planning in the domain of military air campaign planning. The demonstration took place in May and June 1996, for both DARPA and Rome Laboratory representatives and for representatives from Air Combat Command, Langley Air Force Base. The demonstration was well received, and as a result has become a cornerstone of ongoing work under the DARPA/Rome Labs Planning Initiative, the program under which it was funded.

SRI's SIPE-2 planning system was used to perform generative planning. As the above figure shows, SIPE-2 was integrated with three other systems -- ACPT, an air campaign plan authoring tool developed at ISX; CTEM, a force requirements estimator and scheduler developed at AEM Services; and the Plan Visualization Tool (PVT) developed by General Electric and based on its Tachyon system. The role of SIPE-2 was to accept a partial plan created with ACPT; expand the unsolved goals in the plan by using the knowledge encapsulated in SIPE-2's air campaign planning knowledge base (ACP KB) and supplemented by recommendations made by CTEM; and pass the resulting plan to PVT for inspection.

The resulting integrated system provides several capabilities to an Air Force staff planner: a feasibility analysis in terms of resources and time required to execute the plan; a visualization of the plan, including resource and schedule shortfalls, in an easily understandable form; and an environment for plan modification.

The planning problem

The problem confronting an air campaign planner is complex. Given a set of high-level political and military goals (for example, "Protect US citizens and forces from hostile attack"), the planner refines the goals that are attainable (wholly or in part) by the employment of air power into more specific goals. This process iterates until each goal is directly attainable by the execution of a mission. A group of identical aircraft acting in concert performs a mission. Each mission consists of a mission type, a time and place, a type of aircraft, and the number of sorties required to execute the mission. Thus, a mission might be expressed as "Four F-15Cs to escort strike package P to target T on day D+1." Low-level mission planning details, such as flight path and altitude profile, are outside this application's scope.

There are often multiple ways to refine goals into subgoals. These refinements reflect the different strategies and tactics that are available. For example, a refinement of the preceding goal might include a subgoal to defend a friendly country F that is near a belligerent nation. Further refinements might contain defensive tactics ("Patrol the borders of F"), preemptive tactics ("Attack hostile airbase AB near F"), or a combination thereof. Available options are constrained by the situation, which includes local geography, the enemy's characteristics and capabilities, restrictions imposed by political authority, and the availability of aircraft and other assets.

These strategies, tactics, and constraints are represented in the ACP KB, developed in consultation with the Checkmate office of the US Air Force. The knowledge base covers gaining and maintaining air superiority over both friendly and enemy territory, force application against weapons of mass destruction, and support requirements (refueling, reconnaissance, and protection from air and ground threats). The ACP KB also represents the results of the (human-conducted) intelligence analysis of the situation, which serves to focus the generative planner on enemy strengths, weaknesses, and other salient aspects of the situation.

The planning technology

The plans generated for this demonstration are the largest ever generated by SIPE-2, containing several thousand primitive actions. We accomplished this application with only one small change to SIPE-2, purely for efficiency reasons. In a low-level routine, SIPE-2 had previously used an algorithm that was quadratic in the number of choice points in the plan. We replaced this algorithm with a linear algorithm (made possible by sorting the choice points).

The application system

To understand the integrated system and its capabilities, it is useful to follow a plan as it is processed by the components of the system. A human operator interacts with ACPT to create an initial plan. The human planner successively decomposes and refines the plan's high-level political and military goals until they meet one of two criteria: the goal is resolved to the point at which the ACP KB has knowledge about how to solve it ("Achieve air superiority over friendly forces and enemy territory"); or the goal is resolved to a set of targets. The former goals pass to SIPE-2 for further refinement, and the latter pass to CTEM.

When the user selects the option of finishing the plan automatically, all goals are extracted from the ACPT plan, converted from the plan's object-oriented database representation into a CLOS representation, and passed to SIPE-2. SIPE-2 extracts the appropriate attributes from each CLOS goal and builds a corresponding SIPE-2 goal. Goals that meet the first criterion are collected into a partial plan that SIPE-2 expands using the ACP KB.

After resolution, some goals -- all those that do not involve attacking targets -- will have been resolved down to the primitive-action level; these require no further processing until subsequent scheduling and resource allocation. The remaining target goals -- those generated by SIPE-2 as well as by the human planner -- are collected and passed to CTEM. CTEM recommends the type and number of strike assets (either cruise missiles or a combination of aircraft and munitions), and schedules the strikes on the basis of goal priority and availability of strike assets. CTEM also recommends a grouping of strikes into packages of strikes for delivery more or less simultaneously.

Once CTEM has made strike recommendations, the next step is to determine the support requirements of each strike package. These requirements include defense against hostile aircraft and surface-to-air missiles, refueling, and reconnaissance. SIPE-2 generates the corresponding support missions as follows. The CTEM recommendations are input and converted into SIPE-2 support-package goals, which are then collected into a second (post-CTEM) SIPE-2 plan. SIPE-2 then solves these goals, using the ACP KB, which contains knowledge about alternatives for strike package support and the types and numbers of aircraft required.

At this point, all planning goals have been fully expanded into missions. To determine whether the campaign plan is feasible in terms of available resources, all missions are collected from both the pre-CTEM and post-CTEM SIPE-2 plans, and passed to a resource-allocation module. This module allocates resources to missions in a nonoptimized, greedy manner. If resources become exhausted, limited alternatives are explored, and resource shortfalls may be identified.

The completed campaign plan now consists of three components -- the pre-CTEM and post-CTEM SIPE-2 plans, and the human-generated ACPT plan. These are linked by SIPE-2 into a unified view of the campaign plan by associating all goals with their subgoals that reside in another component.

SIPE-2 then writes this unified plan in the input syntax of PVT, for presentation to the human planner. PVT, which replaces the SIPE-2 graphical user interface, was specifically designed for air campaign planners. Missions with resource shortfalls are annotated so that they can be highlighted in the plan display. Furthermore, all goal linkages are conveyed; these linkages enable PVT to propagate resource shortfall information upwards through the goal ancestry, and allow the display of all high-level goals that may fail because of resource shortages detected at lower levels. Mission scheduling information also passes to PVT and propagates upwards. This information is used to display the duration of all goals and missions, and to identify schedule overruns.

PVT highlights problems in the plan. To fix problems, the system operator invokes a plan modification and control module to change the plan. From this moduel, the operator invokes SIPE-2 to expand the modified campaign plan and SIPE-2 displays the revised plan in PVT. Currently, this application supports only the addition of goals and limited modifications of planning assumptions and of available resources. However, SIPE-2 can handle other plan modifications.

In summary, the integrated system provides plan visualization and feasibility estimation, as well as plan modification to fix detected problems.


This effort was sponsored by the Air Force Materiel Command, Rome Laboratory, and the Defense Advanced Research Projects Agency.

SIPE-2 is a trademark of SRI International.

Back to Wilkins Home Page Back to SIPE-2 Home Page

David E. Wilkins
Last modified: Wed Dec 4 12:38:11 1996