Аннотация
This paper addresses the resolution, by Genetic
Programming (GP) methods, of ambiguous inverse
problems, where for a single input, many outputs can be
expected. We propose two approaches to tackle this kind
of many-to-one inversion problems, each of them based
on the estimation, by a team of predictors, of a
probability density of the expected outputs. In the
first one, Stochastic Realisation GP, the predictors
outputs are considered as the realisations of an
unknown random variable which distribution should
approach the expected one. The second one, Mixture
Density GP, directly models the expected distribution
by the mean of a Gaussian mixture model, for which
genetic programming has to find the parameters.
Encouraging results are obtained on four test problems
of different difficulty, exhibiting the interests of
such methods.
Пользователи данного ресурса
Пожалуйста,
войдите в систему, чтобы принять участие в дискуссии (добавить собственные рецензию, или комментарий)