Blackboard systems for knowledge-based signal understanding
Norman Carver, Victor Lesser · Prentice-Hall, Inc eBooks · 1992
A sophisticated signal understanding system requires knowledge-based interpretation techniques as well as sophisticated signal processing techniques. Blackboard systems are the most widely used artificial intelligence (AI) frameworks for understanding and interpretation problems. This paper reviews the blackboard model of problem solving and discusses how blackboard systems can be used for understanding problems. However, it does not specifically show how blackboard systems can be used for knowledge-based signal processing. Integration of signal processing and understanding via blackboard systems is addressed in [35]. We use the term signal understanding rather than signal processing here, because our focus is on the development of “high-level” descriptions of data in terms of the “situations” it represents. Another way of saying this is that signal understanding involves the meaning of the data, while signal processing is generally limited to the extracion of important features—i.e., semantic versus syntactic descriptions. The notion of understanding has also been associated with enabling a system to act on the data in an appropriate manner. Thus, a sound understanding system must do more than just identify harmonically-related acoustic signals, it must also determine their source—e.g., a ringing telephone (that may need to be answered). Likewise, a radar understanding system must do more than just identify echoes, it must also be able to explain the echoes in terms of the movements of certain vehicles, environmental disturbances, and so on. Signal understanding problems cannot typically be solved by applying the kind of algorithmic data transformations used in traditional signal processing. This is because signal understanding applications often involve a great deal of underlying uncertainty: critical features cannot be accurately or precisely “extracted” or with the types of features that can be extracted the space of situations to be considered is still very large. For example, while there may be models of the output that individual sources/targets should ideally elicit from a sensor, it may not be possible to have exact models for the interactions of every possible combination of sources. Add to this the uncertainty over the effect that environmental conditions or other phenomena may have on the sensor output and the problem is very difficult to solve algorithmically. Blackboard systems (and AI techniques in general) are appropriate for understanding problems that involve very large answer spaces as well as uncertain, errorful, or incomplete data and problemsolving knowledge. These characteristics can produce a combinatorial explosion in the number of situation models that must be considered (this issue is pursued further in Section 1). Solving such problems requires an approach that is very different from the algorithmic approaches of traditional signal processing. Instead of attempting to directly determine the best complete answer, in blackboard-based problem solving partial possible solutions are incrementally constructed and