We study the problem of characterizing the behavior of lossy and data corrupting communication channels in a networked control setting, where the channel's behavior exhibits temporal correlation. We propose a behavior characterization mechanism based on a hidden Markov model (HMM). The use of a HMM in this regard presents multiple challenges including dealing with incomplete observation sequences (due to data losses and corruptions) and the lack of a priori information about the model complexity (number of states in the model). We address the first challenges by using the plant state information and history of received/ applied control inputs to fill in the gaps in the observation sequences, and by enhancing the HMM learning algorithm to deal with missing observations . Further, we adopt two model quality criteria for determining behavior model complexity. The contributions of this paper include: (1) an enhanced learning algorithm for refining the HMM model parameters to handle missing observations, and (2) simultaneous use of two welldefined model quality criteria to determine the model complexity. Simulation results demonstrate over 90\% accuracy in predicting the output of a channel at a given time step, when compared to a traditional HMM based model that requires complete knowledge of the model complexity and observation sequence.
ABSTRACTWe study the problem of characterizing the behavior of lossy and data corrupting communication channels in a networked control setting, where the channel's behavior exhibits temporal correlation. We propose a behavior characterization mechanism based on a hidden Markov model (HMM). The use of a HMM in this regard presents multiple challenges including dealing with incomplete observation sequences (due to data losses and corruptions) and the lack of a priori information about the model complexity (number of states in the model). We address the first challenges by using the plant state information and history of received/applied control inputs to fill in the gaps in the observation sequences, and by enhancing the HMM learning algorithm to deal with missing observations . Further, we adopt two model quality criteria for determining behavior model complexity. The contributions of this paper include: (1) an enhanced learning algorithm for refining the HMM model parameters to handle missing observations, and (2) simultaneous use of two well-defined model quality criteria to determine the model complexity. Simulation results demonstrate over 90% accuracy in predicting the output of a channel at a given time step, when compared to a traditional HMM based model that requires complete knowledge of the model complexity and observation sequence.