A Runtime Framework for Adaptive Compositional Modeling

The rapid emergence of embedded devices and sensor networks that frequently exchange object-level images foretells an increasing reliance on object-level systems. Additionally, nearly all computing systems, including control systems, enterprise applications, scientific codes and dynamic libraries operate eventually at the object code level. Studying adaptivity and runtime composition issues in such systems is becoming an important focus of systems research. In this thesis, we describe an object-level framework that will manipulate an object module to instrument control functionality and adaptivity in order to realize complex compositional scenarios. Using function and parameter remapping capabilities, our framework transcends programming language and design boundaries, and enables applications to adapt dynamically during runtime. We introduce the capability to "restart" an application automatically, a feature we utilize to support adaptivity not only spatially, over the algorithm domain, but temporally as well. A high-level adaptive control language based on XML is presented that allows complex adaptive scenarios to be expressed concisely. Additionally, the construction of several adaptive scenarios using our framework is illustrated, along with several experiments in learning adaptivity using reinforcement learning techniques.