Abstract:We consider the identification of non-causal systems with arbitrary switching modes (NCS-ASM), a class of models essential for describing typical power load management and department store inventory dynamics. The simultaneous identification of causal-and-anticausal subsystems, along with the presence of possibly random switching sequences, however, make the overall identification problem particularly challenging. To this end, we develop an expectation-maximization (EM) based system identification technique, where the E-step proposes a modified Kalman filter (KF) to estimate the states and switching sequences of causal-and-anticausal subsystems, while the M-step consists in a switching least-squares algorithm to estimate the parameters of individual subsystems. We establish the main convergence features of the proposed identification procedure, also providing bounds on the parameter estimation errors under mild conditions. Finally, the effectiveness of our identification method is validated through two numerical simulations.
What problem does this paper attempt to address?
### Problems the paper attempts to solve
The paper attempts to solve the identification problem of non - causal systems with arbitrary switching modes (NCS - ASM). Such systems are very crucial in describing typical power load management and department store inventory dynamics, etc. However, simultaneously identifying the causal subsystems and anti - causal subsystems, as well as the possible random switching sequences, makes the overall identification problem particularly complex.
### Main contributions
1. **Proposed for the first time**: As far as the author is aware, this is the first work considering the identification problem of NCS - ASM. In particular, in the E - step of the EM framework, an improved Kalman filter (KF) is proposed to calculate the posterior state estimates of the causal part and the anti - causal part, which is crucial for dealing with the identification of non - causal systems.
2. **Expandability**: Different from the system dynamics in the existing literature where the switching behavior only occurs in the causal part, the method in this paper can handle the switching behaviors in both the causal part and the anti - causal part. Moreover, the switching sequences in the two directions can be different.
### Method overview
The paper adopts the Expectation - Maximization (EM) framework to solve the identification problem of NCS - ASM. The specific steps are as follows:
1. **E - step**:
- Calculate the posterior estimate of the switching sequence using Bayes' rule.
- Use the improved Kalman filter (KF) to estimate the states of the causal part and the anti - causal part.
- Calculate the complete data likelihood function \( Q(\theta, \theta^k) \).
2. **M - step**:
- Maximize \( Q(\theta, \theta^k) \) to update the parameter estimate \( \theta^{k + 1} \).
- Calculate the closed - form solutions of the switching probabilities \( \{\pi_c^i\}_{i = 1}^{m_c} \) and \( \{\pi_a^i\}_{i = 1}^{m_a} \) through the first - order optimality conditions.
- Use the switching least - squares method to calculate the closed - form solutions of the system matrices \( \{A_c^i\}_{i = 1}^{m_c} \), \( \{A_a^i\}_{i = 1}^{m_a} \) and the output matrices \( \{C_c^i\}_{i = 1}^{m_c} \), \( \{C_a^i\}_{i = 1}^{m_a} \).
- Estimate the noise covariance matrices \( \Sigma_c \), \( \Sigma_a \) and \( \Sigma_m \).
### Mathematical model
Consider the following discrete - time, non - causal system with arbitrary switching modes:
\[
\begin{aligned}
& x_c(t)=A_c(s_c(t))x_c(t - 1)+v_c(t), \\
& x_a(t)=A_a(s_a(t))x_a(t + 1)+v_a(t), \\
& y(t)=C_c(s_c(t))x_c(t)+C_a(s_a(t))x_a(t)+v_m(t),
\end{aligned}
\]
where:
- \( t\in\mathbb{Z} \) is the time index,
- \( x_c(t)\in\mathbb{R}^{n_{xc}} \) and \( x_a(t)\in\mathbb{R}^{n_{xa}} \) are the causal state vector and the anti - causal state vector respectively,
- \( y(t)\in\mathbb{R}^{n_y} \) is the system output,
- \( s_c(t)\in\{1, 2,\ldots,m_c\} \) and \( s_a(t)\in\{1, 2,\ldots,m_a\} \) are discrete variables representing possible switching modes,
- \( v_c(t)\in\mathbb{R}^{n_{xc}} \) and \( v_a(t)\in\mathbb{R}^{n_{xa}} \) are system noise vectors,
- \( v_m(t)\in\mathbb{R}^{n_y} \) is the measurement.