Machine learning based link optimization of wireless communications often relies on past experience, accurate estimation of channel conditions, and theoretical performance models. Typically, theoretical models poorly match given situations, past experiences are limited, and spectrum sensing of noise and channel conditions pose many hurdles. Hence, traditional cognitive radio engines based on genetic algorithms and case based reasoning have faltered, especially when facing new environments. Our approach uses efficient experimental designs to generate an empirical performance model as an alternate to theoretical models. The procedure systematically probes the system by setting a unique combination of input parameters and transmitting a data file across the link. Performance metrics, such as packet error rate and throughput, associated with each row of a response surface methodology (RSM) design estimate simple models of performance. Goals of this research include validating accuracy of empirical models based on type and efficiency of experimental design, using empirical models in place of theoretical models during the optimization process, and comparing success of an experimental design driven decision compared to a benchmark genetic algorithm cognitive radio engine. Over-the-air implementation on software defined radios demonstrated the statistical approach performing within 4% of a traditional genetic algorithm cognitive engine even in cases where the statistical fit of the estimation model is poor.