In many circumstances a telephone call can be completed through a connecting network in several ways. Hence, there naturally arise problems of optimal routing, that is, of making the choices of routes so as to achieve extrema of one or more measures of system performance, such as the loss (probability of blocking) or the carried load.
As is customary in traffic theory, a Markov process is used to describe network operation with complete information. The controlled system is described by linear differential equations with the control functions (expressing the routing method, being used) among the coefficients. Restricting attention to asymptotic behavior leads to a problem of maximizing a bilinear form subject to a linear equality constraint whose matrix is itself constrained to lie in a given convex set. An alternative approach first shows that minimizing the loss, and maximizing the fraction of events that are successful attempts to place a call, are equivalent. This fact permits a dynamic programming formulation, which, in turn, leads to a very large linear programming problem. Two small examples are treated numerically by this method.
It is particularly important to try to verbalize, and then mechanize, the optimal routing strategies. In this endeavor, the linear programming formulation is of limited usefulness. Therefore, in the latter half of the work we have attempted to use the special combinatorial structure imposed by the telephonic origins of the problem to shed light on the character of the optimal strategies. In particular, we show that for connecting networks with suitable combinatorial properties, the optimal route choices can be very simply described. Some of the remits obtained were suggested by, and verify, conjectures from the practical lore of telephone routing.
The problem of routing calls falls into two parts: Which attempted calls should be accepted in which states? What route should an accepted call use? The first problem is very hard, and only sample numerical answers for small networks are obtained. We solve the second problem analytically for a large class of cases by appeal to combinatorial structure in the network. These cases can be described roughly as those in which the relative merit of states (as far as blocking is concerned) is consistent or continuous; i.e., if a state x is “better” than another y, then the neighbors of x are in the same sense “better” than the corresponding neighbors of y. An abundance of examples indicates that these cases are numerous and so warrant attention. In a network with this kind of combinatorial property, a policy which rejects no unblocked calls and minimizes the number of additional calls that are blocked by completing an attempted call differs from an optimal policy only in that the latter may reject some calls.