In many domains, such as medical diagnosis, obtaining the complete set of feature values for a test instance is impractical due to the costs associated with the features. Test costs can arise in various forms depending on the problem domain. For instance, in the medical domain, the test costs are typically monetary. Traditional machine learning algorithms focus on the single objective of minimizing error, assuming all features are "free,"' even though they often are not. Learning in the presence of test costs introduces a second objective for the learner to satisfy: minimizing cumulative test costs.;This thesis focuses on solving the dual-objective learning problem for regression using a greedy feature acquisition approach. The approach in this thesis sequentially acquires features of a test instance in the order that is expected to provide the greatest reduction in prediction error per unit cost. This method can be applied to any regression algorithm given a dataset and the test costs associated with each feature. With this approach, greater accuracy can be achieved by acquiring less than the complete set of features for a test instance. Experimental analysis compares the benefits of cost-sensitive attribute acquisition against that of random and cheapest-first selection.