Different multiprocessor interconnection architectures, including the Manhattan Street Network (MSN), have attracted interests as optical packet switching infrastructures. This paper studies the deployment of such architectures as regular virtual topologies in arbitrary physical networks. The inputs to the virtual topology design problem are the physical topology, the traffic matrix and the regular topology. In this paper, this problem is both tackled directly and by decomposition into two subproblems. The first subproblem - dilation minimization - uses only the physical topology and the virtual topology as optimization inputs. Node placement optimization - the second subproblem - considers the traffic matrix and virtual topology as inputs. The solutions of these two subproblems are compared with each other and against the results obtained when the global problem is optimized directly (using all three possible input parameters) for a variety of traffic scenarios. This gives insight into the relative importance of the physical topology and traffic matrix when designing a regular virtual topology for optical packet switching. Regardless of the approach adopted, the problem is intractable, and hence, heuristics must be used to find (near) optimal solutions expeditiously. Five optimization heuristics, using different artificial intelligence (AI) techniques, are employed in this paper. The results obtained by the heuristics for the three alternative design approaches are compared under a variety of traffic scenarios. An important conclusion of this paper is that the traffic matrix plays a less significant role than is conventionally assumed, and only a marginal penalty is incurred by disregarding it in several of the traffic cases considered. In fact, it was found that it is possible to design the regular virtual topology without using the traffic matrix, and yet, the solution is close to optimal for a range of traffic scenarios and relatively immune to traffic fluctuations.