QA And Generalization

> In SBM, Quality Assurance (QA) most often requires some additional, rather laborious actions, like special verification experiments and testing. In EBM QA is a naturally integrated part of the modeling process. And since the examples (used for model training) ARE the reality, this reality IS the model so to say, beside of some measurable errors for the single parameters of the model at hand. Of course, these errors give valuable hints on the suitability of any single parameter, and additionally on the overall modeling quality of a given parameter set, which offers a great help on the way to find the best parameter set for a given problem.

> An immanent problem of EBM is "overfitting", which results in a rather well-fitting model for a given set of examples, but lets the model fail, if more examples (even from the same sources) are added. To avoid this effect, "rotating training and test sets" are used during model training. If there are enough examples available, this is a well-known and solid way to avoid "overfitting" and to ensure GENERALIZATION, to model the true inherent relationships between the parameters. Of course, this procedure can come to some boundaries for "sparse problems", when only a rather small number of examples is available.

> Another general problem for high-dimendionsl models can show up at deployment time: Again because of the max 2-5 dimensional capabilities of human brains, a user can not have a measure, how reliably a model answer may be at a given point in the n-dimensional problem space. In our CAEBM technology, we overcome this problem with the aid of a special neural net, modeling the density of examples over the n-dimensional problem space. So the resulting "reliability indicator" can measure the applicability of the model at hand for any given point in its parameter space.

> In general, CAEBM does not provide a model only for a given problem, but also reliable measures for the model quality, as well in terms of global indicators as for every single parameter involved, which cannot be expected in most SBM processes. And, of course, powerful CA is a most important ingredient to make these CAEBM qualities possible.