Machine understanding of the plan is the key to helping humans deal with the information glut created by advanced situation-awareness systems like SAIM. The plan specifies expectations for how events will unfold, so the EA can compare actual events to the situations that were anticipated. We use rich, knowledge-based plan representations [45] to allow computers to share context with users, so both understand the semantics of plans and requests.
We had two tasks involving significant knowledge acquisition and domain modeling: (1) we had to model SUO plans and the actions that compose them, and (2) we had to model the value of information and various types of alerts for users. We interacted with several domain experts to develop these models. These tasks were aided by the centuries of analysis and modeling that have already been done in this domain. For task 1, the Army already has a standard plan representation called the Operations Order, which has a required structure, but the entries are mostly free text. Primitive actions in this domain are referred to as missions, and there are Army field manuals that describe missions in detail. We modeled missions in a hierarchical mission model. Our mission model and plans are described in Section 6.5. For task 2, there is extensive accumulated experience and analysis of errors and opportunities that arise during execution of SUO plans, but there are many tradeoffs to be made. The tradeoffs and our models are described in Sections 6.4, 6.6, and 6.7.
Mission-specific execution monitoring is achieved by a novel integration of mission knowledge represented as methods with an AI reactive control system. The EA invokes methods at appropriate points during plan execution. The methods employ mission-specific algorithms and in turn invoke EA capabilities in a mission-specific manner. Much of the domain and mission knowledge is encoded in the mission model and not explicitly represented in the plan itself, which specifies a partial order of missions for each team member. The EA uses the plan to invoke the knowledge in the mission model at the appropriate time and with the appropriate arguments.
Another feature of our approach, particularly for terrain reasoning, is the pervasive use of specialized programs, possibly external to the EA, to perform complex computations that are important to system performance. By using alternative specialized programs, the EA can easily adapt the granularity of its reasoning and improve performance as better modules become available. For example, API functions in our design can be used for terrain reasoning and to compute the enemy strength from the current tracks.
Our approach builds on SRI's continuous planning technology [45,47,48] and on the domain-independent Act formalism [46]. Act represents procedural knowledge and plans as Acts, provides a rich set of goal modalities for encoding activity (see Section 6.5), and has been used by several institutions [47,10]. The EA uses PRS [15,48] as its reactive control system (other reactive control systems have similar capabilities, e.g., UM-PRS [10]). PRS is a good framework for constructing the EA because it supports parallel activities within an agent, and can smoothly interleave responses to external requests and events with internal goal-driven activities with its uniform processing of goal- and event-directed behavior. PRS uses procedures encoded as Acts and its extensive graphical tracing provides valuable insights into EA operation.