|
|
Approximate Techniques for Markov Models
Andriushchenko, Roman ; Havlena, Vojtěch (referee) ; Češka, Milan (advisor)
In this work we discuss approximative techniques for the analysis of Markov chains, namely, state space aggregation and truncation. First, we focus on the application of the former method for the analysis of discrete-time models: we redesign the clustering algorithm to handle chains with an arbitrary structure of the state space and, most importantly, we improve upon existing bounds on the approximation error. The developed approach is then integrated with uniformisation techniques, in both standard and adaptive forms, to approximate continuous-time models as well as provide estimates of the approximation error. This theoretical framework along with existing truncation-based techniques were implemented within PRISM model checker. Experiments confirm that newly derived bounds provide a several orders of magnitude precision improvement without degrading performance. We show that the resulting aggregating approach can provide a valid model approximation supplied by adequate approximation error estimates, in both discrete and continuous time. Then, we perform a comparative analysis of aggregating and truncating techniques, illustrate how different methods handle various types of models, and identify chains for which aggregating, or truncating, analysis is preferred. Finally, we demonstrate a successful usage of approximative techniques for model checking Markov chains.
|
|
|
Approximate Techniques for Markov Models
Andriushchenko, Roman ; Havlena, Vojtěch (referee) ; Češka, Milan (advisor)
In this work we discuss approximative techniques for the analysis of Markov chains, namely, state space aggregation and truncation. First, we focus on the application of the former method for the analysis of discrete-time models: we redesign the clustering algorithm to handle chains with an arbitrary structure of the state space and, most importantly, we improve upon existing bounds on the approximation error. The developed approach is then integrated with uniformisation techniques, in both standard and adaptive forms, to approximate continuous-time models as well as provide estimates of the approximation error. This theoretical framework along with existing truncation-based techniques were implemented within PRISM model checker. Experiments confirm that newly derived bounds provide a several orders of magnitude precision improvement without degrading performance. We show that the resulting aggregating approach can provide a valid model approximation supplied by adequate approximation error estimates, in both discrete and continuous time. Then, we perform a comparative analysis of aggregating and truncating techniques, illustrate how different methods handle various types of models, and identify chains for which aggregating, or truncating, analysis is preferred. Finally, we demonstrate a successful usage of approximative techniques for model checking Markov chains.
|
guest ::
