Operational Rationality Through Compilation of Anytime Algorithms
Description:... In standard algorithms, the fixed quality of the output allows for composition to be implemented by a simple call-return mechanism. However, when algorithms have resource allocation as a degree of freedom, there arises the question of how to construct, for example, the optimal composition of two anytime algorithms, one of which feeds its output to the other. This scheduling problem is solved by an off-line compilation process and a run-time monitoring component that together generate a utility maximizing behavior. The crucial meta-level knowledge is kept in the {\em anytime library} in the form of {\em conditional performance profiles}. These profiles characterize the performance of each elementary anytime algorithm as a function of run-time and input quality. The compilation process therefore extends the principles of procedural abstraction and modularity to anytime computation. Its efficiency is significantly improved by using {\em local compilation} that works on a single program structure at a time. Local compilation is proved to yield global optimality for a large set of program structures.
Show description