Optimizing the time warp protocol with learning techniques

The history of the Time Warp protocol, one of the most commonly used synchronization protocols in parallel and distributed simulation, has been a history of optimizations. Usually the optimization problems are solved by creating an analytical model for the simulation system through careful analysis of the behavior of Time Warp. The model is expressed as a closed-form function that maps system state variables to a control parameter. At run-time, the system state variables are measured and then utilized to derive the value for the control parameter. This approach makes the quality of the optimization heavily dependent on how closely the model actually reflects the simulation reality. Because of the simplifications that are necessary to make in the course of creating such models, it is not certain the control strategies are optimal. Furthermore, models based on a specific application cannot be readily adopted for other applications. In this thesis, as an alternative approach, we present a number of model-free direct-search algorithms based on techniques from system control, machine learning, and evolutionary computing, namely, learning automata, reinforcement learning, and genetic algorithms. What these methods have in common is the notion of learning. Unlike the traditional methods used in Time Warp optimization, these learning methods treat the Time Warp simulator as a black box. They start with a set of candidate solutions for the optimization parameter and try to find the best solution through a trial-and-error process: learning automata give a better solution a higher probability to be tried; reinforcement learning keeps a value for each candidate that reflects the candidate's quality; genetic algorithms have a dynamic set of candidates and improves the quality of the set by mimicking the evolutionary process. We describe how some optimization problems in Time Warp can be transformed into a search problem, and how the learning methods can be utilized to directly search for the optimal value for the system control parameter. Compared with the analytical model-based approach, these methods are more generic in nature. Since the search is based on actual run-time performance of different values for the control parameter, the learning methods also better reflect the simulation reality.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.