In statistical learning theory generalisation error is a measure how
well a predictive of a learning algorithm perform on unseen dataset.  The expected error is approximately equivalent
to the empirical error and the error reduces when we increase the dataset. Since
the generalisation of the learning algorithm is done using randomly drawn from
a finite independent and identically distributed samples the model might be
sensitive. One way of overcoming this problem is by relating the
training(given) error to the generalisation error – which possibly would avoid

There are many theories for
generalisation error – each theory has strengths and weaknesses.