Metamodel-based dynamic algorithm configuration using artificial neural networks

We consider the problem of configuring algorithms dynamically by selecting algorithm parameter values adaptively. The research is motivated by the time dependency of system parameters throughout algorithm runtime in servicing systems: Depending on the customer arrival rate, switching algorithm parameters may be advisable to maintain quality of service. To this end, we develop a metamodel-based methodology for dynamic algorithm configuration: We first record algorithm performance under static system parameters. This knowledge is then translated into an artificial neural network (ANN) predicting performance for given system and algorithm parameters. The ANN finally serves as a metamodel determining optimal algorithm parameters dynamically when there is system parameter variation. Overall, the developed generic methodology for dynamic algorithm control facilitates a structured model-based approach to suitably respond to changing system conditions. The outline is adept to practical instantiation as demonstrated in two service systems where control parameters are adjusted adaptively to customer arrival rates.