>The list of OPs (Output Parameters) and IPs (Input Parameters) defines the problem under consideration, and practically any problem (in a network of problems) can be defined this way on about 1 page. The OPs have to be defined in accordance to the answer-requirements of the problem-owner, and the IPs are defined by the questions "What are the important IPs? and "For which IPs can we expect to find values, ie examples?" KHE experts can be helpful to find the most powerful and convenient IPs, together with the problem owner.
> Besides of numeric parameters, nonnumeric parameters often can do a great job to meet the problem-specific know-how situation in a compact and stringent way, especially in highly interdisciplinary contexts, and the resulting mixed parameter sets (MPS) become the key for powerful and successful CAEBM projects. Beside of Neural Nets new modeling techniques like Boosted Decision Tree Forests (BDTF) open new opportunities to use very efficiently (partially) non-numeric parameter sets.
> Additionally, normalization and standardization of the parameters can simplify many problem situations, and let them focus on the problem kernel. And in specific cases taking into account of partial SBM know-how, rule-based know-how and expert know-how from reliable sources can be helpful to concentrate the EBM on the really open part of the problem at hand. All this of course can influence the parameters in the OP + IP list considerably.
> An additional powerful feature of EBM is its capability, to work with "incomplete" sets of IPs: Because of mostly working with SBM, we normally think of having to provide all physically relevant parameters, ie about 500.000 IPs for a car body stiffness model. But in an EBM for the same job it might be sufficient to use eg 20 relevant IPs, which characterise the important differences between the car bodies under consideration. This tremendous reduction of effort is due to the associative, example-based modeling process of EBMs, where "physically complete" IP sets (like in SBM) are not needed.
> Therefore, a carefully, if appropriate iteratively refined list of OPs + IPs, always with the availability of examples (= sets of values for IPs and OPs) in mind, is the important entry point for any successful CAEBM project.