|
|
2002
AAAI Spring Symposium on Information Refinement and Revision for Decision Making:
Modeling for Diagnostics, Prognostics, and PredictionArtificial IntelligenceBBN • Clustering • Trend Analysis • Expert Systems |
|
Home Symposium
Decision
Making in M&D Artificial Intelligence Software
and Data References
Contact
|
Several AI technology areas could enable innovations in information refinement and revision for decision making. We briefly outline some potential connections below. The list is by no means exhaustive and is designed to give just a flavor of the available approaches. To suggest additional technology areas, contact the organizers. BBNBBNs use a probabilistic framework to model dependencies in a diagnostic framework. Extensions use dynamic networks to reflect the need to take into account the changes over time. BBNs can be extended into influence diagrams through the addition of decision nodes which allow modeling of potential actions to take. The price is a considerable increase in computational burden. Clustering techniques are employed in trying to find structure in data such as different failure types. Challenges of these techniques are to be able to deal with high noise content of measurements and with dynamic changes of the process monitored. In trend analysis, system changes are tracked over time and extrapolations allow the prediction of impending failures. Step changes, increase in variance, intermittend changes, etc., can indicate potential problems. Statistical techniques are traditionally found here. Expert Systems have been used for many years to capture heuristic knowledge about system behavior. Diagnostic expert systems allow the identification of problems with user interaction. Uncertainty has successfully been integrated into expert systems but challenges remain regarding updating and maintenance. Neural networks can be used to identify problems on various levels of complexity. In particular where system knowledge is not readily available and the process is known to be stable, neural networks are useful. Neural nets are also popular because of their learning ability which can be used for tuning of system parameters. Hybrid systems try to combine the advantages of several approaches, such as neural fuzzy hybrid systems. A successful tool is ANFIS which combines learning ability with transparency. Similarly, genetic algorithms have been integrated into more traditional tools for optimization purposes. While more recent applications provide some degree of automation, intelligent and semi-autonomous support for creating, searching, and understanding networks will become increasingly important as the size and complexity of data grow. In particular, agents are needed that can filter incoming information and add to an existing network, and that can alert users to interesting events. In changing environments, it is essential for the diagnostic tools to adapt to different external stimuli, but also to different demand profiles, and perhaps to changed system status. This requires advanced learning techniques which no longer make assumptions about stationary systems. Information from several diagnostic sources will only in rare cases agree completely. This is a result of the various underlying models, the data, assumptions made, etc. It is the task of information fusion techniques to resolve potential conflict between the information, taking into account for example historical performance of individual tools. Approaches can be solved for example using Case Based Reasoning Techniques and rule based approaches. Textual passages of fault logs and repair logs are analyzed to extract useful objects (e.g., failure mode, additional observations, etc.) and relations between them (e.g., symptoms observed, repair carried out). Improved techniques for identifying objects and relations are necessary in order to use the vast number of textual documents containing equipment service potential. As a piece of equipment ages and undergoes wear, maintenance actions have
to be carried out which are constrained by the allowable downtime,
availability of maintenance personnel, system status (i.e., emergency repair
vs. routine maintenance), cost, etc. Statistical techniques using Weibull
distributions may find their place here. Case-Based Reasoning (CBR)Many current M&D tasks are approached via prototypical cases. For example, analysts looking for evaluation of real estate point to past successfully identified cases as prototypes – they say "I want to find more cases like this." Case-based reasoning about complex structures will be one essential component for future approaches to equipment service. Decision making techniques will be used which try to optimize over some
type of utility. BBNs can use the capability of decision nodes of influence
diagrams. Other approaches will have to find a suitable substitute using some
form of utility. |
Version 1.0
|
BBN • Clustering • Trend Analysis • Expert Systems Neural Networks • Hybrid Systems • Autonomous Agents |
