`Rapid learning´ techniques for discrete event systems

One of the approaches that have been followed in the study of discrete event systems (DES) is based on the analysis of state trajectories (sample paths) of these systems when analytical techniques fail to provide closed-form solutions or adequate approximations. The theory of perturbation analysis is founded on the fact that observing the behavior of a DES under one set of design or control parameters often allows one to infer its behavior under a different set of such parameters. This strongly suggests the potential for `learning´ about the global behavior of a system by simply observing one sample path. In turn, this implies the potential for adjusting various critical parameters on line so as to continuously improve performance and react to unexpected changes. The term `rapid learning´ is used to capture the idea that this global behavior can be inferred from a single sample path without having to resort to time-consuming and prohibitively costly `trial-and-error´ techniques (e.g. repetitive simulation). The article overviews the rapid learning framework based on stochastic timed automata as models of DES. The `constructability problem´, fundamental to this framework, is formulated, and two specific approaches to its solution are reviewed and compared. Some applications of rapid learning techniques from areas such as manufacturing and communication networks are briefly described. Some developments promising an alternative approach to solving combinatorially hard design and optimization problems are also discussed