3.2.3 Hypothesis Space Traversal
Given a particular search space, defined by the architecture of the network and training data
provided, we could occupy any point in that space to give us a particular hypothesis. How
we choose to traverse the space of possible hypotheses determines what type of ensemble
we will end up with.
Regularisation Methods
Some authors have found benefit from using a regularisation term in the error function
for the network. The justification from this stems from Tikhonov’s work on ill-posed problems [132]. This showed that including another term in the error function, in addition
to the objective function, could effectively control the bias-variance trade-off3. Using this
approach, the error of network i becomes: