Abstract :
In this paper we argue for a particular collaborative architecture for analysis with two primary structures: the Annotated Hypothesis Lattice (AHL), and the Active Coordination Mechanism (ACM). The primary reason for distinguishing these two structures is to clarify the distinct logic involved in each. In the course of collaborative analysis and reasoning, representations are employed with a wide range of intended semantic and pragmatic properties. Some data might represent raw inputs collected in the field, other data, suggestions of what types of evaluation to perform, and yet other, competing interpretations, logs of steps taken, etc. Without some clear semantic and pragmatic model, the distributed activity could easily become chaotic and unproductive. To improve clarity of design, we suggest that two domains be semantically distinguished, formally modeled, and faithfully implemented. The first is the domain of situational hypotheses. Hypotheses are points in a space, structured by a relation of specificity/generality and annotated with evaluations, sources, etc. A collection of points in that space represents the current "state of play" of the analytic process. We refer to this structure as the Annotated Hypothesis Lattice (AHL). In principle, the AHL structure would be monotonic increasing in size, as new hypotheses are added, new relationships uncovered, and new annotations added. However the value of certain annotations may vary over time, as evidence and counter-evidence accrues. The structure would be navigable by users (with appropriate access rights), and certain anchor points, summaries, and so on, would be indexed and made available for direct access or search. Furthermore, the structure would be built out, accretively, via operations performed on it by a distributed collection of independent or cooperating analysts. The second, and distinct, domain is that of the socialbehavioral processes in which the analysts engage. This domain includes mechanisms - or tracking goals and progress made, for triggering actions across the distributed workgroup, and for optimizing across the spectrum of pre-definition, from totally pre-planned, to planned but subject to elaboration and modification, to fully improvised. We refer to this domain as the Active Coordination Mechanism (ACM). Clearly, process descriptions are key to the functioning of the ACM and require data representations, but the import of these representations, unlike those of the AHL, is prescriptive, and they require carefully designed mechanisms to convert them to action while allowing - and even eliciting - improvisation, elaboration, and task-appropriate spontaneous communication. The elements proposed for this part of the architecture include: (1) a coordination engine which manages an "activity graph," containing current tasks (goals, processes) in play, links to analysts in their roles as owner, stakeholder, etc., and other data. The coordination engine also interprets and enforces the social protocols by which new task are created and task parameters are updated; (2) a triggering mechanism ("prompter") which presents to individual analysts those task items requiring their attention and highlights useful operations that might be performed on them; and (3) a library of useful templates - reusable plan fragments or schemas that can be accessed, modified, instantiated, and launched. The ACM is designed to combine the benefits of structured workflows with improvised local action; formal status signals with informal side communications; asynchronous communication with synchronous exchanges; and active engagement of attention with analyst-initiated operations. In the talk we provide illustrations of dynamic, adaptive planning and execution in the ACM as applied to incremental unfolding and evaluation of the hypothesis space. We will also discuss load balancing, embedded automation (interoperating with humans via a shared social protocol), and the role of analytic
