2.2 Ambiguous Situations

The following are examples where the very nature of the problem appears to be ambiguous from the point of view of the engineered problem solution:

1.

High-tech devices such as a photocopying machine have one set of states associated with each `normal' mode of operation, and a completely separate set of states for failure modes. Examples of failure modes included paper exhaustion, paper jams, mechanical part failure, incorrect paper in all of the paper trays, and power interruption during operation.

2.

In certain high-consequence operations involving data transfer, data processing must continue during and after periods of data loss. In a context sensitive situation, the data lost might alter the behavior of subsequent processing steps. In this event, assumptions must be made about the lost data so processing can proceed.

3.

Designs are being considered for locomotives that manage peak energy demands independently from average energy demand. Although accurate systems can be imagined for controlling the locomotive energy, assumptions must be made concerning the actual performance of the system. Reacting to discrepancies between the actual and predicted performance of the locomotive currently involves extremely complicated heuristics. One of the more common concerns of railroad operators is the loss of efficiency due to changes in air pressure and the friction loss between the tracks and the rails as the wheels wear. In the former, it is very difficult to determine what the actual performance of the locomotive is until it is measured, while in the latter the possibility does exist for a trend to be measured.

4.

The performance of tools to translate between different data forms could be greatly improved if they were less ad-hoc and convoluted. A project that one of the authors has worked on involves attempting to match spatial data between systems operating at different scales and resolution. Even the location of the same feature is difficult to identify, given that it may have different representations and actually appear to be situated in different locations in absolute space.

5.

A huge majority of the traffic lights in the US are capable of operating on a network and acting in an adaptive mode. As of December 1996, officials at the California Transportation Department (CALTRANS) estimated that, nation-wide, only about 5% of overall traffic lights were operating in any kind of adaptive mode.

6.

Cases have been found in the work here at Sandia National Laboratories where the answer depends on the order in which the data is received, yet the data order cannot be known ahead of time.

There has been a considerable amount of research put into developing systems capable of dealing with ambiguous situations (at least from the stand-point of handling inconsistent or uncertain data). In general these are referred to as either a non-monotonic logic system, assumption based truth maintenance system, or a directed backtracking grammar. The foundation of this work is covered by de Kleer [6]. Unfortunately, this work is abstract, and the implementations are too inefficient for deployment in high consequence applications. Further, none of the Lisp (or AI implementations, in general) have any method of generating domain specific representations of the ambiguity. Given the number of examples we have identified where there is some form of ambiguity or inconsistency in the information flowing into applications, we conclude there is a need for:

1.

formal computation models based on some form of non-monotonic logic

2.

research into the structure of formal languages for specifying inconsistent and ambiguous data, and

3.

better tools for generating formal models based on (1) and (2).