Computer based learning techniques need a representative sample of
power system situations, for which topology 
, electrical state
and margins 
and 
are pre-determined.
The construction of such samples calls for random sampling techniques
similar to those developed in the decision tree approach to power
system security assessment [8], and exploits an
appropriate ``system-theory'' LPM computation method.
Figure 3: Principle of the data base generation
The overall principle of such a data base generation is depicted in
Fig.
3, where we consider a total number of
different contingencies, 
different candidate attributes, and
an overall sample of 
operating states, where 
denotes the
number of states used in the learning set to derive the approximate
models, and 
the number of independent test states used
subsequently to validate them.
The random sampling approach aims at screening all relevant power system situations, and in particular normal (usual) states as well as weak situations. The power system engineers are generally able to provide valuable prior information helping to determine the considered random variations (e.g. load level, generation schedule, topology, voltage set points ...).