Sliding Hidden Markov Model for Evaluating Discrete Data

摘要

The possibility of handling infrequent, higher density, additional loads, used mainly for on-line characterization of workloads, is considered. This is achieved through a sliding version of a hidden Markov model (SlidHMM). Essentially, a SlidHMM keeps track of processes that change with time and the constant size of the observation set helps reduce the space and time complexity of the Baum-Welch algorithm, which now need only deal with the new observations. Practically, an approximate Baum-Welch algorithm, which is incremental and partly based on the simple moving average technique, is obtained, where new data points are added to an input trace without re-calculating model parameters, whilst simultaneously discarding any outdated points. The success of this technique could cut processing times significantly, making HMMs more efficient and thence synthetic workloads computationally more cost effective. The performance of our SlidHMM is validated in terms of means and standard deviations of observations (e.g. numbers of operations of certain types) taken from the original and synthetic traces.