 |
 |
 |
 |
| |
|||||
|
|
|
Many decision-support applications monitor distributed, heterogeneous databases to assess time-critical decisions. Users of such applications can easily be overwhelmed with data that can change rapidly, conflict, or be redundant. Developers are faced with a dilemma: either filter out most information and risk excluding critical items, or gather possibly irrelevant information and overwhelm the decision maker. This article describes a solution to this dilemma called decision-centric information monitoring (DCIM). DCIM is based on the simple observation that decision-support applications should monitor only information that can potentially change some decision. In our research on Software Agents, MITRE has developed and prototyped an approach to DCIM that meets the requirements implied by this observation. This article presents those requirements and briefly describes our prototype implementation. To see the need for automated assistance, consider a logistics officer supporting a military operation where it may not be feasible to wait until all supplies are in hand before beginning a campaign. Consequently, the logistics plan sequences deliveries such that the resources needed to support each phase of an operations plan are delivered in time to the right location. Therefore, success of the operations plan depends on a complex, interdependent series of logistics movements, the timeliness of which are affected by weather conditions, availability of transport, bottlenecks at ports and air bases, and so on. The logistics officer must continually monitor diverse information sources to assess the viability of his or her plan and the success of the operations plan. This officer needs help determining which of the thousands of potentially relevant conditions needs to be monitored, given the current mission context. Requirements for DCIM A comprehensive approach to DCIM must meet three main requirements: 1) Relevance: Selecting information to monitor should be based on the value of information to the user. Identification of relevant information in large-scale, dynamic applications often is not possible by humans. 2) Robustness in multi-database environments: DCIM must cope with heterogeneity, distribution, autonomy, and intermittent connection of sources as well as redundant, incomplete, and conflicting information. 3) Dynamic Response: Determination of information value should be recomputed upon the receipt of each new piece of information. This requirement is acute for time-critical decisions, where options change as a situation evolves. The information monitored will change depending on the following two conditions: a) The "stage" of the user’s decision process: During planning, the user will want information relevant to generating alternatives and projecting consequences. Once the user has committed to a plan of action, interest shifts to information relevant to changing decisions, including whether projected consequences and preconditions will occur. b) What is already known: Value of information is incremental and context-specific. For example, if one has evidence of a particular occurrence, it may be less important to monitor other sources for confirmation than to search for disconfirming evidence. Prototype Implementation In the Intelligent Software Agents research project we developed an approach to DCIM that meets the requirements above. We implemented a prototype, called LOOKOUT, to investigate implementation issues associated with DCIM. Given a military logistics plan, LOOKOUT monitors distributed, heterogeneous data sources and notifies the user when he should consider replanning. As the figure illustrates, the key features of our approach include the following: Decision Model To focus on information most likely to alter decisions, DCIM requires an executable model of explicit decision criteria to determine what conditions are relevant. A decision model for DCIM must support sensitivity analysis of options available to a decision maker. It is increasingly common to find computer-executable decision models in use in time-critical decision-support applications that meet requirements for DCIM decision modeling. Techniques from Artificial Intelligence, Operations Research, and Decision Analysis are suitable. LOOKOUT used a knowledge-based planning system that generates and monitors the execution of military logistics plans. We developed techniques for transforming such plans into Bayesian networks, which could be used to deduce the probability of plan success. Sensitivity Analysis Sensitivity analysis is necessary to identify what information has potential to change decisions. The product of sensitivity analysis is a list of conditions to which decisions are sensitive. We call these items the user’s critical information needs. We invoked the Bayesian network to identify changes in information that would substantially reduce the probability of plan success. Those changes defined the critical information needs. Mediation Usually, the critical information items identified by sensitivity analysis do not exactly match specific data items in available sources. When there is such a discrepancy, one must translate between the two levels. Mediation is the process of developing mappings to enable this translation. Often, one must cope with a "semantic gap" between more abstract concepts in the decision model and more concrete ones in the sources. Information Monitoring Information monitoring converts specific data requests into queries, triggers, and other implementation artifacts that can be executed by component data sources. The requests specify the data to be retrieved, required timeliness, quality of service, and other parameters. Our research led to a new approach to information monitoring that exploits commercial data replication technology. As a result of conducting this research, MITRE gained considerable knowledge about database query optimization techniques. By applying this knowledge to the Office of Naval Intelligence’s Seawatch database, we were able to improve the performance of frequently executed multimedia queries by a factor of 12. Feedback When information monitoring gathers relevant information from data sources, mediation is invoked to aggregate and map data into concepts used by the decision model. This transformed decision-relevant data is sent to a decision maker and to the model. Using the updated data, sensitivity analysis is performed again, yielding a possibly different set of information needs. These are translated to a new set of retrieval requests. In this way, the monitoring process continually provides information relevant to the current decision problem. User notification in LOOKOUT takes place if critical troops or equipment are unlikely to arrive at a destination within specified time constraints. Notification takes place via pop-up alerts, email, or alphanumeric pager. Discussion Lessons learned from developing LOOKOUT include the following: First, we found that previous research approaches to situation monitoring did not apply to our customers, because they did not sufficiently respect the autonomy of participating data sources. This led to the development of our "smart caching" approach that uses commercial replication technology. Another key lesson was the existence of a semantic gap between decision model concepts and those typically represented in real world information sources. For example, sensitivity analysis against the decision model might reveal that Loading Facility Status is a critical information item. However, there is probably not a single database item that indicates this. Instead, Loading Facility Status might be a compilation of several data items about numbers and condition of fork lifts, availability of personnel, facilities and supplies to support them, etc. While complete automation is beyond the reach of today’s technology, the development of semi-automated tools for bridging the semantic gap is a promising area for future research. Lessons learned from this effort have fed a number of efforts including DARPA’s Battlefield Awareness and Data Dissemination program and its Agile Information Control Environment program. In addition, our experiences provided much of the inspiration for DARPA’s Plan Sentinel concept, an important requirement for the Advanced Logistics Project. A plan sentinel determines what conditions and information sources should be monitored based on analysis of a plan, and then spawns agents that perform the monitoring. Our software provided the first demonstration of this concept. For more information, please contact Len Seligman using the employee directory. |
Solutions That Make a Difference.® |
|
|
