An expert systems approach to decision support in a time-dependent, data sampling environment

An approach to the use of raw data that greatly reduces the volume of data while increasing its value for analysis and decision-making is considered. Consistent application of an analyst's knowledge in an automated environment readily transforms raw data into information and the information into recommendations. The approach concentrates first on capturing the analyst's knowledge. The method permits the incremental, and fairly random, capture of knowledge at several levels of detail. The human view of the knowledge base is transformed into the computer view which is then applied to the raw data.<<ETX>>


INTRODUCTION
In order to make recommendations about a system or to determine how a system reacted during a given period of time, analysts must process a huge volume of raw data. The sheer volume of data can be staggering, especially in fields such as electronic warfare systema testing. The resulting difficulty in sifting useful information from the data leaves little time for coming to meaningful conclusions. The analyst needs a consiatent method of applying knowledge about a system. This method should work well in an automated environment which converts raw data into meaningful in f or mat i on and recommendations. The intent is to quickly determine what a system did during a given period of time and make recommendations accordingly.
The flow of data from its r a w form to suggested recommendations is illustrated in Figure 1. As the data move through this scenario, they become lower in volume but higher in value relative to the goal of the analysis. Figure 1 illustrates three transformation phases: raw data collection, rtate analysis, and recommendation determination. The computer must pass the raw data through the state analyris and recommendation phaaea and deliver appropriate report8 uaing the knowledge supplied by the analyst. Time period8 which contain no active atater and later, no offered recommendationr, can be deleted from consideration. Each of the three phases will now be di8cusred.

RAW DATA COLLECTION
The collection and organization of raw data is a crucial step in any automated analysis. Data from each system in the test must be gathered, recorded, converted, and passed on to the state analysis phase (Figure 2).

MRi] = Recommendation Made
In Cell if_(Offered Or Not) T L

STATE ANALYSIS
Two things must happen to support the state analysis phase. First, the knowledge of the analyst must be captured. Tha supporting mechanism must permit incremental, and fairly random, knowledge capture. Second, this knowledge must be ueed in a consistent manner, along with the raw data, to determine what the system did during any given period of time.
Capturing the Analyst's Knowledge. The knowledge of the analyst must be gathered and organized into a detailed description of how ryrtem states are recognized in the raw data. Based on extensive interviews with experienced analyata, three basic categories make up this knowledge base: Accept-Only-If Rules: rules which must be satisfied before a state, event, or set is determined to exist. These rules are tied to a particular state, event, or set and are invoked only i f the given state, event, or set is indicated in the raw data. Acceptance is not accomplished unless all the rules arm satisfied.

States
time an event must last. Where Qx can be a state name or a specification set.

An example of a Reject-Only-If rule is:
I f BAD-TIME then Reject. Meanr that the state, event, or set that the rule is tied to will be determined to not exirt only i f the time stamp ir incorrect.

An example of an Accept-Only-If rule ia:
If SUB-SYSTEM-B then Accept.
Meanr that Sub-System-B must be active before the indicated Btate, event, or ret can be meaningfully idantified.
It is very important that an analyatoriented mechanism be devised for the capture of these seven categories of knowledge. The mechanism chosen must uae the language of the analyst and must be eimple enough for the analyst to incrementally add knowledge in a fairly random fashion. One mechanism for doing this is the expanded spreadeheet shown in Figure 3.  The Multi-Dimensional Network Database Representing the State Analysie Knowledge Base Tha network ehown in Figure 4 is traversed in a forward-chaining, breadth-f irat fashion. The inference engine ueee the network to acquire the information it neede to tell when given eete, evente, and rtater occur.

STATE NAME
The travereal of tho network ie euah that eoah #ample of data il compared to the specification #eta of the current event of each #tote on the recommendation is associated with a liet of criteria which muet be eatiefied before that recommendation can be suggeeted. Each criteria has an aeociated 'history' factor. This factor is a period of poet time counting from the preeent time over which the given criteria may have been active. For instance, it may be important to know whether or not a switch wae in the 'On' position during the past five seconds.
The 'history' factor can be specified as greater than zero i f history is important or made equal to zero i f only the present condition of the criteria is needed. A mechanism for capturing and organizing this type of knowledge can take the form of a relational database's input screen and the database management software's editing, deleting, and inserting functiona. Figure 5 illustrates the input screen. The tree shown in Figure 6 is traversed in a top-down, depth-first fashion. The left branch of each node is the branch taken i f the recommendation or state making up the given criteria is currently active. The right side is taken i f the recommendation or state is not currently active. This process continues until a suggested kecommendation is achieved.

RECOMMENDATIONl: recommendation name
Then, the recommendation report and the active recommendation list are updated. The process continues down the right branch of the node until the leave8 of the tree are reached. Each time a recommendation or state is added to or removed from the active list, the binary recommendation tree must be traversed. In a previous paper, Raeth 2 i Hardin discussed their partial implementation and use of this recommendation process.

R3
C

SUMMARY
Thie papor hae diecueeed a method of using raw data and a n analyat'e knowledge in an automatod environment to turn data into information and information into suggested rocommendatione.
[For a good tutorial on oxpert ryetoms in general, eee Gevarter.3 The contribution# made by thie paper are: 1) a way of getting meaning out of raw data samples.
2) a way of helping the analyet cope with tho growing mounds of r a w data while prorerving tho detail of the data.
Tho curront diroction of thie project i a to document the aoftware already produced for state analyeie and to move that software into full production.