Stochastic MCLP: Audio & Facility Location

• Stochastic MCLP is a unified framework that models multi-sensor sequential data using probabilistic methods and adaptive Kalman filtering.
• It integrates joint spatial-temporal filtering and maximum likelihood estimation to enhance speech dereverberation and suppress interference.
• In operations research, it underpins robust facility location models with scenario-based optimization, capturing uncertainty and cooperative effects.

Stochastic Multi-Channel Linear Prediction (MCLP) defines a unified and principled statistical framework for processing multi-sensor sequential data, with its principal applications in speech dereverberation, interference suppression, and stochastic optimal facility location under uncertainty. The concept spans both signal processing (notably in multi-microphone audio enhancement) and operations research (notably in location-optimization problems under uncertainty and cooperation). Stochastic MCLP models late reverberation, noise, and signal structure as random processes, leveraging maximum likelihood and/or Bayesian filtering (e.g., Kalman filtering) to adapt models online in dynamic or uncertain environments. In location optimization, stochastic MCLP describes a scenario-based, multiperiod, and cooperative demand covering model, capturing uncertainty in facility attraction and leveraging large-scale mixed integer optimization.

1. Stochastic MCLP in Signal Processing: Probabilistic Model Formulation

In the multi-microphone speech enhancement context, stochastic MCLP treats the desired early-reflection speech signal as a time-varying complex Gaussian random variable, typically within the short-time Fourier transform (STFT) domain. For a reference microphone signal $d_1[n, k]$ (frame $n$, frequency bin $k$),

$$d_1[n, k] \sim \mathcal{CN}\left(0, \gamma_{n, k}\right)$$

where $\gamma_{n, k}$ is the unknown, time-varying power spectral density. Independence is assumed across frames and frequency bins, and directional interferers are incorporated as part of residual noise or handled in subsequent spatial filtering stages. This stochastic modeling enables closed-form maximum likelihood (ML) updating of prediction and filtering coefficients, naturally facilitating dynamic (time-varying) scenarios (Chetupalli et al., 2019).

2. State-Space Models and Kalman Filtering for MCLP Adaptation

Time-varying environments necessitate adaptive models. In the dynamic framework, the late-reverberation predictor coefficients $G_n$ or their stacked representation are modeled as a linear-Gaussian state-space process:

$$g_{m,n} = g_{m, n-1} + i_{m, n},\quad i_{m, n} \sim \mathcal{CN}(0, \Lambda_n)$$

where $g_{m, n}$ are the MCLP predictor coefficients and $\Lambda_n$ is the innovation covariance. The observation equation is

$$x_m[n] \sim \mathcal{CN}\left(g_{m, n}^H \phi[n], \gamma_n\right)$$

with $\phi[n]$ stacking lagged microphones. The Kalman filter provides recursive prediction and update equations, producing real-time tracking of late-reverberation statistics and filter adaptation. This approach is empirically validated for both stationary and dynamic/moving sources (Chetupalli et al., 2019). In the ISCLP framework, spatial (generalized sidelobe cancellation, GSC) and temporal (MCLP) filters are stacked into a joint state vector $w(l)$:

$$w(l) = \begin{pmatrix} w_{\text{SC}}(l) \ w_{\text{LP}}(l) \end{pmatrix} \in \mathbb{C}^{M L - N_T}$$

and evolved via

$w(l) = A^H(l) w(l-1) + \Delta w(l)$

where $A(l) = \sqrt{\alpha} I$ and $\Delta w(l)$ is zero-mean process noise. The resulting Kalman recursion minimizes a-posteriori error variance and supports low-complexity, $O(M^2)$, adaptation per frame (Dietzen et al., 2019).

3. Joint Spatial-Temporal Filtering and Maximum Likelihood Estimation

Stochastic MCLP frameworks, such as RTF-MCLP and ISCLP, tightly couple spatial filtering (e.g., minimum variance distortionless response, MVDR) with temporal prediction. The observation model in the STFT domain is

$\mathbf{x}[n] = \mathbf{a} d_1[n] + G^H \phi[n]$

where $\mathbf{a}$ is the relative transfer function vector. The joint likelihood is optimized via block-coordinate descent:

1. Late-reverb filter $G$ update (weighted prediction error):

$$\hat{G} = \left(\sum_n \gamma_n^{-1} \phi[n] \phi^H[n]\right)^{-1} \left(\sum_n \gamma_n^{-1} \phi[n] \mathbf{x}^H[n]\right)$$

1. RTF vector update via empirical covariance.
2. MVDR spatial filter update:

$$\hat{\mathbf{w}} = \frac{R_{\hat{d}\hat{d}}^{-1} \hat{\mathbf{a}}} {\hat{\mathbf{a}}^H R_{\hat{d}\hat{d}}^{-1} \hat{\mathbf{a}}}$$

1. Variance updating as instantaneous residual energy.

These steps are iterated to convergence (typically $3$–$5$ cycles). For the dynamic case, all updates are integrated into a filtering recursion per frame (Chetupalli et al., 2019, Dietzen et al., 2019).

4. Stochastic MCLP in Combinatorial Optimization: Facility Location Under Uncertainty

In operations research, the stochastic Maximal Covering Location Problem (MCLP) addresses optimal facility placement under uncertain demand-capture. A planner chooses, over periods $T$, which facilities (sites $J$, types $K_j$) to install/upgrade, constrained by per-period budgets $b^t$. Customer demand attraction $a_{ijk}^{ts}$ is scenario-dependent ($s \in S$). The objective is to maximize the expectation (over uniformly weighted scenarios) of covered demand, with class weights $n_i^t$:

$$\max \sum_{t \in T}\sum_{i \in I} n_i^t \frac{1}{|S|} \sum_{s \in S} z_i^{ts}$$

where $z_i^{ts}$ is binary coverage of class $i$ in $(t, s)$. Partial attractions aggregate via an Ordered Median function, permitting flexible modeling of cooperative effects:

$$U_i^{ts} = \sum_{r=1}^{|J|} \lambda_{ir} u_{i(r)}^{ts}$$

with $u_{i(r)}^{ts}$ the $r$-th largest attraction and $\boldsymbol{\lambda}_i$ nonincreasing weights. The demand is covered only if $U_i^{ts} \geq T_i^{ts}$ (Domínguez et al., 2023).

5. Exact Optimization Algorithms and Valid Inequality Techniques

The embedded assignment problem for the Ordered Median function creates a nonlinear, bilinear model. Stochastic MCLP is solved by:

• Mixed-Integer Linear Programming (MILP) reformulation: bilinear terms $u \cdot \sigma$ are replaced by auxiliary variables $w_{ijr}^{ts}$, with McCormick bounds and scenario-indexed constraints.
• Valid inequalities that tighten the LP relaxation, such as

$\sum_r w_{ijr}^{ts} \leq u_{ij}^{ts}$

• Generalized Benders Decomposition: master problem in $(x, z)$ controls facility/opening variables, while scenario-level assignment subproblems are solved in $O(|J| \log |J|)$ via closed-form primal–dual recursions. Benders cuts exploit dual supergradient vectors for fast convergence (Domínguez et al., 2023).

This toolbox enhances computational tractability, especially for large-scale, multiperiod real instances.

6. Empirical Results and Performance Analysis

Computational studies compare baseline MILP (SL), MILP with valid inequalities (VI), and Benders decomposition (B):

• On synthetic data ($|T| = 3$, $|I|$ up to $50$, $|J|$ up to $30$, $|S|$ up to $10$): SL solves approximately 35% of instances to optimality in one hour, VI 60%, and B 70%. When not solved, Benders' mean MIP gaps are $<10$\%, SL/VI often exceed 30%.
• On real-world data (EV-station location in Trois-Rivières, $|I|$ up to $317$, $|J|$ up to $30$), Benders solves up to $80\%$ of instances (smaller $|J|$), with median times of hundreds of seconds. For large $|J|$, solution gaps remain under $\sim 15\%$.
• Regret analysis: using a non-cooperative $\lambda = (1, 0, \dots)$ covering model when clients actually cooperate worsens demand coverage by 1–3%; layouts change significantly across $\lambda$-vectors, substantiating the value of modeling cooperation via Ordered Median.

In all cases, introducing valid inequalities reduces solution time, but Benders decomposition delivers superior scalability and robustness (Domínguez et al., 2023).

7. Significance and Synthesis Across Domains

Stochastic MCLP provides a rigorous statistical-and-optimization framework applicable both to real-time adaptive signal processing and to scenario-based combinatorial optimization. In speech enhancement, it unifies WPE/MCLP dereverberation and MVDR/beamforming under ML and state-space models, enabling robust operation as environments change (Chetupalli et al., 2019, Dietzen et al., 2019). In facility location, stochastic MCLP generalizes classical and cooperative coverage models, supports flexible objective aggregation (Ordered Median), and yields new algorithmic strategies for robust planning under uncertainty (Domínguez et al., 2023). A plausible implication is that stochastic MCLP—by leveraging scenario-based modeling, convex relaxations, and recursive estimation—can serve as a template for other fields wherein uncertainty and cooperation fundamentally affect optimal action.