Meta Models to Aid Planning of Intelligent Machines
RAND GRADUATE SCHOOL SANTA MONICA CA
Pagination or Media Count:
Relatively simple low-resolution models are needed by human planners and probably by intelligent machines. Ideally, these should be high-level models developed in a multiresolution, multiperspective modeling MRMPM framework. That, however, is often difficult. We ask whether statistical meta modeling i.e., development of response surfaces can provide good low-resolution models if one already has a credible higher-resolution base model. We ask how meta models compare if they are derived from pure statistical methods, from a phenomenology-rich theoretical approach, or from some synthesis. To sharpen issues and generate insights, we have worked through a particular problem in detail. Our conclusions are generally negative about purist statistical meta models, which have serious shortcomings in explanatory power, in variance, and in ability to predict and explain the relative importance of contributing variables. Purely theoretical approaches, however, are often very difficult and not transparent. Fortunately, a synthesis of methods is feasible and likely to be fruitful. Some tentative principles are that 1 a thoughtful first-order theoretical analysis conducted with MRMPM principles in mind can identify aggregation fragments to be used as variables in generalized regression and 2 this can also suggest structures to impose on the meta model that will assure dependences known to be important. Imposing such a structure can, e.g., assure that a meta model will predict failure of a system if any of its critical components fail. The theory-enhanced statistical meta model may also be much better than a naive statistical meta model in representing a systems performance when a competitor is systematically looking for a circumstances that will defeat the system.
- Computer Programming and Software