Neural Network Control Framework

• Neural network control frameworks integrate NN architectures into controller policies, system models, and embedded optimization for dynamic system control.
• They leverage supervised, unsupervised, and physics-informed learning techniques, enabling robust performance in safety-critical applications such as robotics and industrial systems.
• Methodologies include differentiable optimization layers, neural surrogates, and real-time MPC adaptations that improve control precision and computational efficiency.

A neural network based control framework is an integrated methodology wherein neural networks are used as explicit parameterizations of controller policies, system models, or surrogate optimization problems within feedback loops for dynamical systems. These frameworks span supervised, unsupervised, and physics-informed learning paradigms, are compatible with both model-based (e.g., MPC) and direct closed-loop architectures, and encompass not only standard feedforward networks but also advanced differentiable optimization layers and recurrent architectures. Their application domain includes safety-critical mechatronics, robotics, industrial plants, vehicular systems, and beyond.

1. Mathematical Structure and Control Problem Formulation

Neural network control frameworks typically solve the following classes of optimal control problem:

• Finite-horizon constrained optimal control (MPC):

$$\min_{\{u_k\}_{k=0}^{N-1}}\,J(x_0,\{u_k\}) = x_N^\top Q_N x_N + \sum_{k=0}^{N-1} (x_k^\top Q x_k + u_k^\top R u_k)$$

subject to

$$x_{k+1} = f(x_k, u_k),\quad x_k \in \mathcal{X},\,u_k \in \mathcal{U},\,x_0=\bar{x}$$

• Continuous-time control via neural ODEs:

$\dot{x}(t) = f(x(t), \mathrm{NN}_\theta(x(t), t))$

with cost

$$J(\theta) = \mathbb{E}\left[\int_0^T \ell(x(t), u(t)) dt + V(x(T))\right]$$

The control law is parameterized by a neural network, either as

• direct state-feedback $u_k = \mathrm{NN}_\theta(x_k)$, or
• solution to a differentiable quadratic program with an embedded neural network approximator for dynamics and/or cost (Winqvist et al., 2020).

Hybrid configurations include joint identification-control learning, differentiable physics-informed neural surrogate models, and embedded optimality conditions (2503.06995, Böttcher, 6 Oct 2025).

2. Neural Surrogate Modeling and Data Generation

Neural surrogates are constructed by training neural networks to approximate either the system dynamics or the optimal control law. Several methodologies are utilized, tailored to the plant and control objective:

• Supervised learning using offline-generated data: Training inputs are generated via efficient sampling schemes such as hit-and-run within the maximal control-invariant set $\mathcal{C}_\infty$, and labels via offline solution of the MPC optimization (e.g., via OSQP or CVXPY layers) (Winqvist et al., 2020).
• Physics-informed neural networks (PINNs): The loss function features both empirical state-transition errors and explicit physics residuals, e.g. $$L(\theta) = \lambda_\text{data} MSE_\text{data} + \lambda_\text{phy} MSE_\text{phy}$$, where the physics term enforces consistency with identified payload effects and system dynamics (2503.06995).
• Self-supervised identification-control dual learning: Models are trained end-to-end to jointly fit system transitions and enforce closed-loop Lyapunov stability via constructed vector fields (Saha et al., 2020).
• Unsupervised, principle-driven learning: In certain frameworks, no expert control samples are needed; the network is directly trained to satisfy Pontryagin Minimum Principle (PMP) or Hamilton-Jacobi-Bellman (HJB) optimality, e.g. via a neural co-state predictor and PMP-informed loss (Lian et al., 16 Jul 2025, Onken et al., 2020).

Data Sampling Table

Method	Sampling	Label Generation
Hit-and-run	Uniform in $\mathcal{C}_\infty$	OSQP / MPC QP
PINN	Simulation data, phys. params	Physics model + empirical
Lyapunov-based	Random, safe ball and iterated expansion	System transitions

3. Neural Network Architectures and Differentiable Layers

Architectural choices are governed by task constraints, control objectives, and desired properties:

• FC ReLU networks (Black-box NN): Pure function approximators for direct control mapping (Winqvist et al., 2020).
• Projection-augmented NN (ProjNN): Enforce hard input/set constraints by adding a differentiable projection onto $\mathcal{U}$ as the final layer (Winqvist et al., 2020).
• OptNet (QP-layer NN): Embed a differentiable QP as an optimization layer, parameterized by features (e.g. $q(x), b_{eq}(x)$), allowing joint learning of cost and constraints (Winqvist et al., 2020).
• LSTM and RNN-based controllers: For temporal dependencies or multi-agent settings, stacked LSTM layers enable memory-efficient feedback mapping (Alsalehi et al., 2021, Liu et al., 30 Sep 2025, Huang et al., 2022, Gordon et al., 2023).
• Physics-informed MLPs: Multi-layer networks with physics inputs, e.g. mass parameters, as direct inputs (2503.06995).
• Co-state predictor networks: For PMP-based regulation, the NN outputs co-state trajectories which are then used to extract optimal control via a QP solve (Lian et al., 16 Jul 2025).
• Distributed GNN controllers: Graph neural networks propagating local features for topologically-varying systems such as power converters (Jakobeit et al., 10 Jan 2026).

4. Integrated Evaluation Metrics and Robustness Analysis

Framework-specific validation and comparison metrics are essential for quantitative assessment:

• Mean Squared Error (MSE): $$\mathrm{MSE} = \frac{1}{N} \sum_{i=1}^N \|\hat u(x_i) - u^*(x_i)\|^2$$ between NN controller and the true optimal law (Winqvist et al., 2020).
• Worst-case constraint violation (WCV): $\mathrm{WCV} = \max_i \max\{0, G \hat u(x_i) - h\}$ (Winqvist et al., 2020).
• Closed-loop performance loss (CLPL): Gap in accumulated cost between NN controller and true optimum, computed over T-step rollouts (Winqvist et al., 2020).
• Reachability and verification tools: Lipschitzian over-approximation of closed-loop reachable sets, scalable to high-dimensional problems (Zhang et al., 2023, Goyal et al., 2020).
• Tracking and robustness metrics: Root-mean-square error (RMSE) across temporal/spatial state trajectories, constraint violation statistics, and quantiles under process/measurement noise (Huang et al., 2022, 2503.06995, Gordon et al., 2023).
• System-theoretic stability criteria: Local exponential stability certified via LMIs for bilinear neural feedback (Shah et al., 30 May 2025), Lyapunov analysis or uniform ultimate boundedness in adaptive/robust controllers (Li, 16 Jun 2025, Saha et al., 2020).

5. Implementation Strategies and Computational Aspects

Implementation spans both supervised offline training and real-time online inference:

• Training: Batch gradient descent or Adam/RMSProp for parameter estimation; for embedded QP-layers, implicit differentiation via KKT conditions handles backpropagation (Winqvist et al., 2020).
• Deployment: Millisecond inference via GPU or ARM CPU acceleration; e.g., PINN and OSQP rolling out in $<6$ms per NMPC step at 100Hz rates (2503.06995), acados NMPC with DNN surrogate in $<1.4$ms (Gordon et al., 2023).
• Event-triggered adaptation: Online optimization of network parameters is activated in response to prediction errors, leveraging buffer history for robustness (Li, 16 Jun 2025).
• Online MPC/NMPC loop: At each step, measure the current state, predict or optimize the next control input via the neural surrogate/QP, propagate dynamics, repeat.

6. Extensions: Multi-Agent, Image Feedback, Graph-Based, and Neuromorphic Control

Neural control frameworks have been extended in several directions:

• Multi-agent, spatio-temporal logic control: LSTM-based neural policies trained on PSO+SQP-optimized data, supporting satisfaction of complex spatial temporal requirements (Alsalehi et al., 2021).
• Socially aware traffic flow regulation: LSTM controllers with Social Value Orientation (SVO) to balance egoistic and collective traffic objectives; utility-based objective functions tunable via SVO angle $\phi$ (Liu et al., 30 Sep 2025).
• Image-based feedback control: Dual-network DQN with Error-Driven Learning (EDL), robustly converging in high-dimensional pixel space, and validated against classical RL algorithms (Raghavan et al., 2021).
• Graph neural networks for generalized interconnection topology: Meta-control policies amortized over converter distributions, leveraging node and edge feature propagation, with distributed switch control heads (Jakobeit et al., 10 Jan 2026).
• Neuromorphic and spiking neural network frameworks: Membrane potential encoding, predictive coding firing rules, and hierarchical CNS-inspired module architectures for estimation and control—achieving hardware efficient and robust concurrent control (Ahmadvand et al., 2023, Pang et al., 6 Nov 2025).

7. Summary of Observed Tradeoffs and Best Practices

Empirical evidence from the above frameworks yield several design and training tradeoffs:

• Projection-augmented and OptNet architectures guarantee feasibility at modest parameter cost, but increase per-step computation time (Winqvist et al., 2020).
• Training sample size beyond a threshold offers diminishing returns; efficient sampling (hit-and-run) is preferred for scalable coverage (Winqvist et al., 2020).
• Physical consistency and explicit identification (PINN + payload adaptation) yield superior tracking compared to fixed-model NMPC (2503.06995).
• Shrinking zone and bias-corrected fusion networks in agro-hydrological ZMPC mitigate model mismatch at increased economic cost (Huang et al., 2022).
• Neural co-state/PMP frameworks require no expert controls and outperform MPC in smoothness and runtime by orders of magnitude, but lack general formal stability guarantees (Lian et al., 16 Jul 2025, Onken et al., 2020).
• Bilinear neural feedback controllers can be certified robust by LMI analysis constructed offline, compatible with real-time online iteration (Shah et al., 30 May 2025).
• Ensemble and modular architectures (multi-agent, SNN, distributed GNN) expand applicability and support topological or biological heterogeneity (Alsalehi et al., 2021, Pang et al., 6 Nov 2025, Jakobeit et al., 10 Jan 2026).

These practices collectively enable systematic, data-driven, and structure-aware design, training, and verification of neural network based control frameworks across diverse application domains and system types.