Autonomic computing systems must be able to detect and respond to errant behavior or changing conditions with little or no human intervention. Clearly, decision making is a critical issue in such systems, which must learn how and when to invoke corrective actions based on past experience. This paper describes the design, implementation, and evaluation of MESO, a pattern classifier designed to support online, incremental learning and decision making in autonomic systems. A novel feature of MESO is its use of small agglomerative clusters, called sensitivity spheres, that aggregate similar training samples. Sensitivity spheres are partitioned into sets during the construction of a memory-efficient hierarchical data structure. This structure facilitates data compression, which is important to many autonomic systems. Results are presented demonstrating that MESO achieves high accuracy while enabling rapid incremental training and classification. A case study is described in which MESO enables a mobile computing application to learn, by imitation, user preferences for balancing wireless network packet loss and bandwidth consumption. Once trained, the application can autonomously adjust error control parameters as needed while the user roams about a wireless cell.