This paper studies practical limitations of learning methods for resource management in non-stationary radio environment. We propose two learning models carefully designed to support rate maximization objective under user mobility. We study the effects of practical systems such as latency and reliability on the rate maximization with deep learning models. For common testing in the non-stationary environment we present a generic dataset generation method to benchmark across different learning models versus traditional optimal resource management solutions. Our results indicate that learning models have practical challenges related to training limiting their applications. The models need environment-specific design to reach the accuracy of an optimal algorithm.