Nowadays, scientists, researchers, and practical engineers face a previously unseen explosion of the richness and the complexity of problems to be solved. Besides the spatial and temporal complexity, common tasks usually involve non-negligible uncertainty or even lack of information, strict requirements concerning the timing, continuity, robustness, and reliability of outputs, and further expectations like adaptivity and capability of handling atypical and crisis situations efficiently.
Model based computing plays important role in achieving these goals, because it means the integration of the available knowledge about the problem at hand into the procedure to be executed in a proper form, acting as an active component during the operation. Unfortunately classical modeling methods often fail to meet the requirements of robustness, flexibility, adaptivity, learning, and generalizing abilities. Even soft computing based models may fail to be effective enough because of their high (exponentially increasing) complexity. To satisfy the time, resource and data constraints associated with a given task, hybrid methods and new approaches are needed for the modeling, evaluation, and interpretation of the problems and results. A possible solution is offered to the above challenges by the combination of soft computing techniques with novel approaches of anytime and situational modeling and operation.
Anytime processing is very flexible with respect to the available input information, computational power, and time. It is able to generalize previous input information and to provide short response time if the required reaction time is significantly shortened due to failures or an alarm appearing in the modeled system; or if one has to make decisions before sufficient information arrives or the processing can be completed. The aim of the technique is to ensure continuous operation in case of (dynamically) changing circumstances and to provide optimal overall performance for the whole system. In case of a temporal shortage of computational power and/or loss of some data, the actual operation is continued maintaining the overall performance “at lower price”, i.e., information processing based on algorithms and/or models of simpler complexity provide outputs of acceptable quality to continue the operation of the complete system. The accuracy of the processing may become temporarily lower but it is possibly still enough to produce data for qualitative evaluations and supporting further decisions.
Situational modeling has been designed for the modeling and control of complex systems where the traditional cybernetics models haven’t proved to be sufficient because the characterization of the system is incomplete or ambiguous due to unique, dynamically changing, and unforeseen situations. Typical cases are the alarm situations, structural failures, starting and stopping of plants, etc. The goal of situational modeling is to handle the contradiction arising from the existence of a large number of situations and the limited number of processing strategies, by grouping the possible situations into a treatable (finite) number of model classes of operational situations and by assigning certain processing algorithms to the defined processing regimes. This technique - similarly to anytime processing - offers a tradeoff between resource (including time and data) consumption and output quality.
The presentation gives an overview of the basics of anytime and situational approaches. Besides summarizing theoretical results and pointing out the arising open questions (e.g. accuracy measures, data interpretation, transients), the author enlightens some possibilities offered by these new techniques by showing successful applications taken from the fields of signal and image processing, control and fault diagnosis of plants, analysis and expert systems. Some of the discussed topics are:
- Anytime Fuzzy Fast Fourier Transformation and Adaptive Anytime Fuzzy Fast Fourier Transformation: How can we determine the most important signal parameters before the signal period arrives? How can we implement fast algorithms with only negligible delay?
- Anytime Recursive Overcomplete Signal Representations: How can we minimize the channel capacity necessary for transmitting certain amount of information? How can we provide optimal and flexible on-going signal representations, on-going signal segmentations into stationary intervals, and on-going feature extractions for immediate utilization in data transmission, communication, diagnostics, or other applications if the transmission channel is overloaded and in the case of processing non-stationary signals when complete signal representations can be used only with serious limitations?
- High Dynamic Range (HDR) imaging and situational image quality improvement: How can we make the invisible details of images visible? How can we enhance the useful information of images which is significant from the point of view of further processing?
- Anytime control and fault diagnosis of plants: How can we produce useful results and react in crisis situations very quickly in order to avoid catastrophes? How can we increase the safely available reaction time of the (slow) human supervisor by significantly decreasing the time needed for the automatic detection and diagnosis of faults?
- CASY, an Intelligent Car Crash Analysis System: How can we build an intelligent expert system, capable to reconstruct the 3D model of crashed cars autonomously (without any human interaction) using only 2D photos; based on it, how can it determine characteristic features of crashes like the energy absorbed by the car-body deformation, the direction of impact and the pre-crash speed of the car? In what other fields can the algorithms of system be used?