Data augmentation is becoming essential for improving regression accuracy in critical applications including manufacturing and finance. Existing techniques for data augmentation largely focus on classification tasks and do not readily apply to regression tasks. In particular, the recent Mixup techniques for classification rely on the key assumption that linearity holds among training examples, which is reasonable if the label space is discrete, but has limitations when the label space is continuous as in regression. We show that mixing examples that either have a large data or label distance may have an increasingly-negative effect on model performance. Hence, we use the stricter assumption that linearity only holds within certain data or label distances for regression where the degree may vary by each example. We then propose MixRL, a data augmentation meta learning framework for regression that learns for each example how many nearest neighbors it should be mixed with for the best model performance using a small validation set. MixRL achieves these objectives using Monte Carlo policy gradient reinforcement learning. Our experiments conducted both on synthetic and real datasets show that MixRL significantly outperforms state-of-the-art data augmentation baselines. MixRL can also be integrated with other classification Mixup techniques for better results.