Steven Finette
Acoustics Division
Naval Research Laboratory
Washington DC 20375
steven.finette@nrl.navy.mil
Lay-language version of paper 3pID2
Presented on Wednesday, November 12, 2008
156th Meeting of the Acoustical Society of America
As organizations and governments rely more and more on making decisions based on the results of complex computer models, it is important to develop methods that allow for uncertainty to be taken into account in the modeling process itself. In this manner, model predictions will more faithfully reflect our knowledge of the reliability of information incorporated in the model.
verybody makes decisions in situations where the information available to them is sometimes incomplete or uncertain. For example, you get into your car and drive 4 miles to the supermarket to pick up the bottle of milk you forgot yesterday during your weekly food shopping trip. In doing so, you assessed that the risk of having an accident during the extra trip was so small that it would not deter you from pursuing your goal. The decision you made was based on information and experience accumulated over some period of time. The probability (i.e., plausibility or likelihood) of having an accident depends on the weighting of the available evidence, and you were comfortable enough with your risk assessment to perform the task of driving to the supermarket and back home. Decisions like this one are routinely made all the time by individuals, often without much thought. Roughly speaking, one internally constructs a mental “model” of the situation, feeds in some relevant information that is considered reliable and makes a decision based on the output (the likelihood of having an accident). You made a prediction that it was very unlikely you would have a car accident, based on weighting the available information.
An analogous series of events should occur when evaluating the predictions made by computer models(often called computer simulations) of natural phenomena. These models represent a form of complex scientific hypothesis in which the model is specified by mathematical statements about the physical system under study and then translated into a language that a computer understands. The computer then solves the equations, and outputs information corresponding to a prediction. One important question that arises is this: How confident is the scientist in the prediction that the model produces and how does he quantify his impression of the significance of the result?
As computer models become more and more sophisticated, they often describe systems of greater and greater complexity and this question becomes increasingly difficult to answer. In ocean acoustics, for example, one studies how sound propagates and gets distorted in an ocean environment. The distortion introduced by the ocean environment makes communicating in the ocean via sound waves much harder than communicating by cell phones using electromagnetic waves in air. Remote sensing of sound emitted by acoustic sources underwater is relevant to many applications, like determining the location of the sound source, identifying the nature of the source as well as communication between a source and a receiver.
To model the propagation of sound, a scientist has to specify both the ocean environment and characteristics of the acoustic source in terms of a set of parameters and fields. Such input data include the sound speed field, i.e. the value of the speed of sound in the water at each point between the source and receiver, as well as the depth of the ocean, roughness characteristics of the ocean surface and ocean bottom, density of matter in the bottom, etc. Acoustic characteristics such as source depth and frequency must also be specified. All this information provides necessary conditions for obtaining a solution to the problem of simulating the propagation of sound through the ocean. Once this information is supplied, a computer can solve the appropriate equations. However, if the information supplied to the model is incomplete, then the information is not sufficient and the resulting prediction has to be considered suspect. In modeling sound propagation in the ocean, incomplete knowledge of these quantities is the rule rather than the exception and, therefore, there is a need to improve our ability to make computer based predictions in the presence of uncertain information.
Researchers are attempting to include uncertainty within the numerical simulation framework in order to improve the predictive capability of computer modeling. Some of the methods used involve studying the sensitivity of the acoustic field to changes in the environmental parameters. In this way, one could focus in on the most important quantities or combination of quantities that describe the distortion of the acoustic wave in a particular environment, and place special emphasis on trying to decrease the uncertainty in those quantities by making better measurements or possibly constructing more robust models. Other approaches treat the parameters as probabilistic quantities and try to incorporate this uncertain information directly into the simulation itself. In this way, the prediction also has a measure of uncertainty built into it and gives the researcher an ability to make a prediction in the presence of incomplete knowledge. A goal of the research in this subject is to enhance the predictability of computer simulation in complex situations where one doesn’t necessarily have the ability to compare the results to experiment.