Genetic algorithms and call admission to telecommunications networks

We consider the admission problem for the two class M/M/C/C queueing system. The admission policy is coded as a binary string and strings which produce good policies are found using genetic algorithms. We use three policy coding methods: Direct codes: The policy consists of a sequence of 2-bit words, corresponding to the admit/deny decision for each class, associated with each state. The resulting binary string represents the policy. Block Coding: A policy string is constructed from a code book of binary strings and a string which dictates the sequence in which code words appear in the policy string. The genetic algorithm operates on both the code book and the sequence string. Program Codes: A binary string is supplied as the program for a simple Turing machine to which the system state is supplied as input. The output of the program is the admission policy for that state. This method is used for very large problems where the direct or block codes are prohibitively long. All three are found to work well for smaller problems and suggested structural properties for good policies. Results were compared to optimal policies derived using a Markov decision process. For very large problems, the program codes easily produced policies with performance comparable to the best known reservation policies.