Cognitive-Flexible Deep Stochastic Model

• The paper demonstrates how integrating deep stochastic latent representation with a bounded cognitive flexibility mechanism achieves safe and rapid online adaptation.
• It employs a surprise-driven online update mechanism within a Bayesian MPC framework to maintain probabilistic safety under abrupt system and observation shifts.
• Empirical evaluations show CF–DeepSSSM outperforms nominal and robust MPC, ensuring optimal safety, comfort, and bounded parameter drift in nonstationary environments.

The Cognitive-Flexible Deep Stochastic State-Space Model (CF–DeepSSSM) is a framework for online adaptation and safety-certified control of partially observed, nonstationary systems. It combines deep stochastic latent-state modeling, surprise-regulated reorganization of inference mappings (cognitive flexibility), and a Bayesian Model Predictive Control (BMPC) architecture with adaptive constraint tightening. The approach is motivated by the challenge of learning-enabled control under abrupt changes in system dynamics or sensing, where fixed-representation models often fail to preserve safety and performance in the presence of distributional shift (Nuchkrua et al., 31 Jan 2026).

1. Stochastic State-Space Modeling and Latent Representation

CF–DeepSSSM models partially observed systems with dynamics and observations: $$x_{t+1} = f(x_t, u_t, w_t), \quad o_t = h(x_t, v_t)$$ where $x_t \in \mathbb{R}^n$ (true state), $u_t \in \mathbb{R}^m$ (input), $o_t \in \mathbb{R}^p$ (observation), $w_t$ and $v_t$ (unknown disturbances).

A deep stochastic state-space model for the latent belief $z_t$ is constructed via: $z_t \sim q_{\phi_t}(z_t \mid o_{0:t}, u_{0:t-1})$ where $q_{\phi_t}$ is a variational inference network parameterized by $\phi_t$. The latent transition and observation dynamics are defined as: $x_t \in \mathbb{R}^n$0 with $x_t \in \mathbb{R}^n$1 and $x_t \in \mathbb{R}^n$2.

The control objective is formulated in belief space, minimizing expected stage costs under chance constraints: $x_t \in \mathbb{R}^n$3 for $x_t \in \mathbb{R}^n$4, enforcing safety (Nuchkrua et al., 31 Jan 2026).

2. Cognitive Flexibility Index and Bounded Reorganization

To moderate adaptation of the inference mechanism, the Cognitive Flexibility Index (CFI) is defined to regulate the reorganization of the latent belief mapping $x_t \in \mathbb{R}^n$5. The CFI constraint enforces bounded change in inference parameters: $x_t \in \mathbb{R}^n$6 Operationally, updates to $x_t \in \mathbb{R}^n$7 are limited such that $x_t \in \mathbb{R}^n$8. This prevents uncontrolled drift of internal representations during adaptation, providing a mechanism for safe and localized model reorganization (Nuchkrua et al., 31 Jan 2026).

3. Online Surprise-Driven Adaptation Mechanism

Online adaptation in CF–DeepSSSM is driven by predictive surprise, which quantifies the negative log-likelihood of an observation under the current model: $x_t \in \mathbb{R}^n$9 At each time step, the following procedure is executed:

1. Inference: $u_t \in \mathbb{R}^m$0, with $u_t \in \mathbb{R}^m$1 the observation and input history.
2. Control: Solve the BMPC using the current belief.
3. Action & Observation: Apply $u_t \in \mathbb{R}^m$2, record $u_t \in \mathbb{R}^m$3.
4. Surprise Calculation: Compute $u_t \in \mathbb{R}^m$4.
5. Parameter Update: Perform a stochastic gradient step for the generative parameters,

$u_t \in \mathbb{R}^m$5

where the step-size $u_t \in \mathbb{R}^m$6 is chosen to satisfy $u_t \in \mathbb{R}^m$7.

This surprise-driven adaptation enables rapid recovery and bounded reorganization in response to abrupt model mismatches, while ensuring that safety is not compromised (Nuchkrua et al., 31 Jan 2026).

4. Embedding in Bayesian Model Predictive Control

CF–DeepSSSM integrates its adaptive latent state model within a finite-horizon Bayesian Model Predictive Control (BMPC) framework: $u_t \in \mathbb{R}^m$8

$u_t \in \mathbb{R}^m$9

To account for epistemic uncertainty in the model, an adaptive constraint tightening procedure replaces each nominal constraint $o_t \in \mathbb{R}^p$0 with: $o_t \in \mathbb{R}^p$1 where $o_t \in \mathbb{R}^p$2 is the predictive mean and $o_t \in \mathbb{R}^p$3 is the state covariance. Under Lipschitz-continuity, this guarantees probabilistic satisfaction of the original constraints. The first input is applied; the problem is re-solved at the next step, ensuring recursive feasibility and probabilistic safety at all times (Nuchkrua et al., 31 Jan 2026).

5. Theoretical Guarantees

CF–DeepSSSM provides several central theoretical guarantees under standard (regularity, Lipschitz, bounded noise, compactness) assumptions:

• Bounded Posterior Drift (Theorem 1): For parameter perturbations $o_t \in \mathbb{R}^p$4 and step-size $o_t \in \mathbb{R}^p$5,

$o_t \in \mathbb{R}^p$6

• Recursive Feasibility (Theorem 2): Adaptive constraint tightening ensures that feasibility at time $o_t \in \mathbb{R}^p$7 implies feasibility at $o_t \in \mathbb{R}^p$8.
• Closed-Loop Input-to-State Stability (Theorem 3): The closed-loop belief dynamics are ISS with respect to bounded modeling error.
• Probabilistic Safety Preservation (Corollary): All $o_t \in \mathbb{R}^p$9 pairs satisfy $w_t$0 with prescribed violation probabilities at every $w_t$1.
• Dominant Tightening (Lemma 1): If $w_t$2 is $w_t$3-Lipschitz and $w_t$4, enforcing $w_t$5 yields

$w_t$6

These results provide strong assurances for stability and safety in nonstationary belief dynamics under abrupt or gradual changes (Nuchkrua et al., 31 Jan 2026).

6. Empirical Evaluation and Comparative Analysis

Simulation experiments are conducted on a two-dimensional partially observed system subject to state and input constraints ($w_t$7, $w_t$8). The stage cost penalizes tracking error on $w_t$9 and regulates $v_t$0 to zero.

Two main scenarios are considered:

• Abrupt Dynamics Shift ($v_t$1):
  • The true system dynamics $v_t$2 switch regime.
  • CF–DeepSSSM exhibits a spike in predictive surprise ($v_t$3), a temporally localized rise in $v_t$4, followed by rapid convergence.
  • State/input constraints remain satisfied throughout.
  • Compared to nominal MPC, which violates safety ($v_t$5), and robust MPC, which achieves perfect safety at higher cost, CF–DeepSSSM maintains perfect safety ($v_t$6), with optimal comfort cost ($v_t$7) and bounded mean CFI ($v_t$8).

Controller	SafetyRate	ComfortCost	Mean CFI
Nominal MPC	0.87	0.92	0.05
Robust MPC	1.00	1.18	0.04
CF–DeepSSSM	1.00	0.78	0.17

• Observation Drift ($v_t$9):
  • System matrices $z_t$0 constant; $z_t$1 (observation mapping) smoothly drifts (sensor miscalibration).
  • CF–DeepSSSM adapts the observation mapping via bounded cognitive flexibility, maintaining high tracking performance and continuous safety (Nuchkrua et al., 31 Jan 2026).

7. Synthesis and Significance

CF–DeepSSSM synthesizes deep stochastic modeling, surprise-constrained latent space reorganization, and rigorous Bayesian MPC for safe online adaptation. Its key attributes include:

• Dynamic adaptation of the inference mapping under formal CFI bounds.
• Explicit embedding in a probabilistically safe BMPC, with adaptive tightening reflecting epistemic uncertainty.
• Theoretical guarantees of bounded parameter drift, recursive feasibility, closed-loop ISS, and probabilistic safety.
• Superior empirical performance compared to nominal and robust MPC under system or sensing shift scenarios.

The methodology demonstrates that learning-enabled, as opposed to learning-based, control architectures can offer both safety and rapid adaptation in nonstationary, partially observed cyber-physical systems (Nuchkrua et al., 31 Jan 2026).