The structural properties of a composite laminate can be tailored by allowing the fibres to vary their orientation within the individual plies. This class of composites is termed variable stiffness and they are able to enhance the structural performance over traditional laminates by redistributing the internal load of the structure. Therefore, it is of interest to pinpoint the structural properties that may benefit most from this concept and the extent thereof. To that end, a multi-objective optimisation of a variable stiffness cylinder was performed which identifies the Pareto front for objectives such as buckling, natural frequency, stiffness and strength. Manufacturing considerations were also incorporated by accounting for the allowable tow curvature and the resulting overlaps during the process. The stiffness variability was formed by means of a Lagrange polynomial, the non-linearity of which is determined by a specified amount of control points. An investigation on the suitable degree of non-linearity of the functional representation of the cylinder was thus carried out. In order to complete the optimisation in a practical time frame, machine learning algorithms were utilised to act as surrogate models and perform predictions on the optimisation objectives. Different algorithms were implemented for that task, namely Radial Basis Function, Kriging and Artificial Neural Network models. The Artificial Neural Network surrogate model outperformed the others when the dimensionality of the problem was high, although the opposite was observed for low dimensional problems. Cylinders with and without cut-outs were optimised separately and compared with their optimised constant stiffness counterparts. It is found that the performance gains for the cylinder without cut-outs are not significant when the additional mass due to the overlaps is accounted for. On the other hand, the cut-out cylinder greatly benefits from the stiffness variation for multiple objectives. Additional structural improvements were witnessed for higher degrees of non-linearity of the utilised Lagrange polynomial.