We study a continuous version of the capacity and flow assignment problem
(CFA) where the design cost is combined with an average delay measure
to yield a non convex objective function coupled with multicommodity flow
constraints. A separable convexification of each arc cost function is proposed
to obtain approximate feasible solutions within easily computable gaps from
optimality. On the other hand, DC (difference of convex functions) programming can be used
to compute accurate upper bounds and reduce the gap.
The technique is shown to be effective when topology is assumed
fixed and capacity expansion on some arcs is considered.