White-Box Neural Ensemble for Vehicular Plasticity: Quantifying the Efficiency Cost of Symbolic Auditability in Adaptive NMPC

TL;DR

Problem: enable vehicular plasticity—adaptation to varying regimes—while preserving formal auditability for safety-critical control. Approach: a white-box NMPC built from a Modular Sovereignty ensemble of eight frozen neural specialists whose dynamics are blended as $\dot{\mathbf{x}} = \sum_{i=1}^{N} w_i \Psi_i(\mathbf{x},\mathbf{u})$, with weights constrained by $\sum_i w_i = 1$ and estimated by a Governor on a sliding window, all evaluated on a fully symbolic CasADi graph. Key contributions: integration of NMPC with physics-informed, frozen specialists; explicit symbolic Jacobians; quantification of the symbolic traversal cost (approximately a 72x–102x slowdown) and rapid adaptation (~7.3 ms); and functional validation under friction and fleet heterogeneity achieving sub-0.1 m position error with a hybrid training regime. Significance: establishes a verifiable, transparent path to adaptive vehicular control under safety standards, while outlining concrete optimization strategies (pruning, asynchrony, stability guarantees, and semantic interpretability) to bridge the gap to real-time deployment.

Abstract

We present a white-box adaptive NMPC architecture that resolves vehicular plasticity (adaptation to varying operating regimes without retraining) by arbitrating among frozen, regime-specific neural specialists using a Modular Sovereignty paradigm. The ensemble dynamics are maintained as a fully traversable symbolic graph in CasADi, enabling maximal runtime auditability. Synchronous simulation validates rapid adaptation (~7.3 ms) and near-ideal tracking fidelity under compound regime shifts (friction, mass, drag) where non-adaptive baselines fail. Empirical benchmarking quantifies the transparency cost: symbolic graph maintenance increases solver latency by 72-102X versus compiled parametric physics models, establishing the efficiency price of strict white-box implementation.

Figures (1)

• Figure 1: Well-Trained PINN performance, Hybrid training ($\text{RMSE}\approx4\times10^{-6}$) closes gap to ODE ceiling. Position error $<0.1$ m, recovery in 0.8 s.