Evolving femtocell algorithms with dynamic and stationary training scenarios

We analyse the impact of dynamic training scenarios when evolving algorithms for femtocells, which are low power, low-cost, user-deployed cellular base stations. Performance is benchmarked against an alternative stationary training strategy where all scenarios are presented to each individual in the evolving population during each fitness evaluation. In the dynamic setup, different training scenarios are gradually exposed to the population over successive generations. The results show that the solutions evolved using the stationary training scenarios have the best out-of-sample performance. Moreover, the use of a grammar which produces discrete changes to the pilot power generate better solutions on the training and out-of-sample scenarios.