On the Design of a Master-Worker Adaptive Algorithm Selection Framework

Overview & Synopsis

We investigate the design of a master-worker schemes for adaptive algorithm selection with the following two-fold goal: (i) choose accurately from a given portfolio a set of operators to be executed in parallel, and consequently (ii) take full advantage of the compute power offered by the underlying distributed environment. In fact, it is still an open issue to design online distributed strategies that are able to optimally assign operators to parallel compute resources when distributively solving a given optimization problem. In our proposed framework, we adopt a reward-based perspective and investigate at what extent the average or maximum rewards collected at the master from the workers are appropriate. Moreover, we investigate the design of both homogeneous and heterogeneous scheme. Our comprehensive experimental study, conducted through a simulation-based methodology and using a recently proposed benchmark family for adaptive algorithm selection, reveals the accuracy of the proposed framework while providing new insights on the performance of distributed adaptive optimization algorithms.