On subscription admission control for network service provision

We consider a novel multiclass subscription admission control network service problem, where the decision maker has complete knowledge of duration and resource requirement of incoming subscription requests. We model it as discounted MDP problem, and prove structural results. In a given state, if it is optimal to admit a certain class, then it is also optimal to admit a more or equally profitable class in a corresponding state. However the optimal policy does not exhibit a structure leading to a simpler implementation of the policy and we numerically solve the finite-state discounted dynamic programming equation using policy iteration algorithm for a two-class problem. We compare the optimal policy with approximate greedy and threshold policies, which are simpler to implement. Computational results for small sizes of the problem suggest that they perform very close to the optimal policy.