Structural risk minimization(SRM) is a best learning algorithm selectionprocedure – choosing the model that minimizes VC upper bound on risk. ,For fixed datasets, as the complexity of the learning algorithm C(Hi)increases, the training error (eˆrr(c)) will decrease but fails to generalize for unseen dataset.
On the other hand, as the VC dimensionincreases, will increase since this depends on the VCdimension. SRC chooses a best model from the sequence of hypothesis ( which minimizes the right-hand side of theabove inequality. As the complex models might over fit and at the same time the moretraining set we have the lower empirical error will be. SRC resolves the trade-offby providing a measure of characterization between complexity and empiricalerror – bybalancing the model’s complexity and training set (empirical) error.