Combined Learning-and-Control

• Combined Learning-and-Control (CLC) is an umbrella term for architectures that integrate adaptive learning and robust control, emphasizing real-time feedback and performance optimization.
• It encompasses diverse methods such as closed-loop stabilization, dual control with GP-based uncertainty, and hierarchical planning that blend learning dynamics with control synthesis.
• CLC techniques leverage model learning, predictive variance, and robust invariant constraints to balance adaptability with guarantees on stability and safety in dynamic environments.

to=arxiv_search.search 天天中彩票微信 human_code to=arxiv_search.search ചികിതරණ 天天中彩票提现 code Combined Learning-and-Control (CLC) denotes a set of research programs in which learning and control are coupled within one optimization, estimation, or feedback architecture rather than being treated as strictly separate stages. In the cited literature, the term covers several distinct constructions: closed-loop feedback added to learning dynamics in generative adversarial networks, dual control with Gaussian-process uncertainty and explicit information gain, separation of offline controller synthesis from online learning of an information state, hierarchical combinations of classical planning with MPC or RMPC and reinforcement learning, robust-control pipelines in which Gaussian-process confidence sets are translated into IQCs, and control-oriented identification in which model learning is regularized by robust control invariant sets (Xu et al., 2019, Alpcan, 2011, Malikopoulos, 2023, Kungurtsev et al., 6 Jul 2025, Fiedler et al., 2021, Mulagaleti et al., 2024). Taken together, these usages suggest that CLC is best understood as an umbrella term for methods in which learning changes the controller or controller synthesis, while control structure constrains, stabilizes, or interprets the learning process.

1. Terminology and scope

Across the literature, CLC is not a single standardized formalism. In some works it means literal closed-loop control of a learning dynamical system; in others it means dual control, bilevel learning-and-control, robust synthesis with learned uncertainty sets, or control-oriented model identification. A plausible implication is that the common denominator is architectural coupling, not a unique mathematical template.

Formulation	Characteristic mechanism	Representative paper
Closed-loop stabilization of learning dynamics	Proportional feedback on discriminator output, implemented as a squared $L2$ regularizer	(Xu et al., 2019)
Dual control with active learning	Weighted trade-off between tracking and GP variance or entropy reduction	(Alpcan, 2011)
Separation-based linear-system CLC	Offline nominal control plus online learning of an information state $\Pi_t$	(Malikopoulos, 2023)
LQR proxy-with-mismatch CLC	Stage-specific mismatch penalties $\beta_t$ learned from data	(Kounatidis et al., 30 Sep 2025)
Mission-aligned hierarchical CLC	Classical planning, MPC/RMPC, and RL in a two-level optimization	(Kungurtsev et al., 6 Jul 2025)
Learning-enhanced robust control	GP confidence sets embedded in LFR/IQC synthesis	(Fiedler et al., 2021)
Control-oriented qLPV identification	Learning regularized by the size of a robust control invariant set	(Mulagaleti et al., 2024)

The breadth of the term is reinforced by additional formulations. Discrete-time Chen-Fliess methods combine model-based control with online learning of residual input-output behavior (Gray et al., 2019). AntLer chooses control parameters by minimizing expected closed-loop cost while explicitly anticipating how future online learning will reduce model uncertainty (Capone et al., 2020). A channel-based framework has also been proposed to unify classical and quantum learning-control loops, treating the evaluator, controller, plant, and switch as interconnected channels (Vedaie et al., 2023).

2. Recurrent architectural patterns

One recurring pattern is feedback injected into the learning dynamics themselves. In GAN stabilization, the adversarial game is treated as a dynamical system, and CLC adds a proportional feedback term $-\lambda D$ to the discriminator dynamics. In Dirac GAN this shifts poles from $s=\pm i$ to the roots of $s^2+\lambda s+1$, changing oscillatory dynamics into a Hurwitz-stable closed loop; in normal GANs the same mechanism is implemented as a squared $L2$ regularizer on discriminator outputs (Xu et al., 2019).

A second pattern is dual control, where control actions are chosen to regulate the system while also improving future identification. In the Gaussian-process formulation, the controller uses predictive means for exploitation and predictive variances for exploration, with entropy reduction as the information metric. The resulting objective is explicitly multi-objective rather than certainty-equivalent: immediate tracking competes with information gathering (Alpcan, 2011).

A third pattern is separation between offline control design and online learning of sufficient statistics. In linear systems, one can design a separated policy offline on a nominal model and then learn online the information state

$\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$

which compresses the growing data stream without loss of optimality. A related LQR-oriented formulation replaces explicit model identification by a proxy objective on the nominal system together with mismatch penalties $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$, and then learns the stage-specific $\beta_t$ directly from data (Malikopoulos, 2023, Kounatidis et al., 30 Sep 2025).

A fourth pattern is hierarchy. In mission-aligned autonomous systems, classical planning operates on symbolic task structure, MPC or RMPC enforces physical feasibility and safety at the motion layer, and RL updates cost models, transition models, or MPC parameters. In robotics, a related hybridization appears in whole-body loco-manipulation, where a model-based admittance controller governs the manipulator, an RL policy controls the legged base, and a Reference Governor enforces constraint satisfaction (Kungurtsev et al., 6 Jul 2025, Schperberg et al., 2 Mar 2026).

A fifth pattern is learning under explicit robust-control structure. One strand learns an unknown component with a GP, translates high-probability confidence sets into sector bounds, and then performs LFR/IQC synthesis over the resulting uncertainty set. Another learns a qLPV model and observer while regularizing the identification objective by the optimal value of a convex quadratic program that certifies existence of a robust control invariant set. In both cases, learning is shaped by what robust control can certify after learning (Fiedler et al., 2021, Mulagaleti et al., 2024).

3. Mathematical formulations

The mathematical expressions used under the CLC label differ substantially by subfield. In GAN stabilization, the controlled transfer function is written in standard closed-loop form,

$\Pi_t$0

with $\Pi_t$1. For Dirac GAN the discriminator and generator transfer functions have poles at $\Pi_t$2 before feedback, while CLC changes the denominator to $\Pi_t$3. In function space for normal GANs, the discriminator flow becomes

$\Pi_t$4

which is implemented by subtracting $\Pi_t$5 or its sample-based analogue from the discriminator objective (Xu et al., 2019).

In GP dual control, information is quantified through Gaussian entropy,

$\Pi_t$6

and the controller solves the myopic weighted objective

$\Pi_t$7

where $\Pi_t$8 is the GP predictive variance. Exact entropy-reduction criteria are derived in terms of log-determinants, while a practical approximation chooses the candidate point of maximum predictive variance (Alpcan, 2011).

In separation-based linear CLC, the sufficient statistic is the information state $\Pi_t$9, and the control law is restricted to separated strategies $\beta_t$0. In the LQR specialization with unknown dynamics, the proxy cost uses stage-specific mismatch penalties,

$\beta_t$1

with dynamic programming on the nominal model and a lightweight outer loop that updates $\beta_t$2 by gradient descent on the realized real-system cost (Malikopoulos, 2023, Kounatidis et al., 30 Sep 2025).

In hierarchical mission-aligned CLC, the high-level policy $\beta_t$3 and learned parameters $\beta_t$4 solve a bilevel program,

$\beta_t$5

where $\beta_t$6 aggregates symbolic and physical states, and the lower problem is an action-triggered MPC or RMPC optimization. In a distinct co-design strand, plant parameters $\beta_t$7 and feedback control are optimized jointly through the HJB value function $\beta_t$8,

$\beta_t$9

with the optimal feedback law

$-\lambda D$0

and $-\lambda D$1 approximated by a PINN that enforces the HJB residual and boundary condition (Kungurtsev et al., 6 Jul 2025, Nishimura et al., 2024).

Other mathematical realizations are equally structured. Chen-Fliess CLC represents the plant input-output map by a discrete-time Fliess operator $-\lambda D$2 and learns residual coefficients online with recursive least squares while retaining model-based predictive control (Gray et al., 2019). AntLer models unknown dynamics with a GP and minimizes a sample-average approximation of expected finite-horizon cost over controller parameters, explicitly simulating how the posterior mean and variance evolve under future online learning (Capone et al., 2020).

4. Representative domains and empirical demonstrations

The term CLC has been used in applications ranging from machine learning itself to robotics, nonlinear control, and robust process control. In GAN training, CLC was evaluated on CIFAR-10 and CelebA with ResNet backbones, Inception Score on CIFAR-10, and FID on CIFAR-10 and CelebA. Reported results include CIFAR-10 IS up to $-\lambda D$3 for hinge loss with spectral normalization, CIFAR-10 FID around $-\lambda D$4–$-\lambda D$5 for CLC-WGAN versus Reg-GAN around $-\lambda D$6 and WGAN-GP around $-\lambda D$7, and iteration speed on CelebA of approximately $-\lambda D$8 it/s for CLC versus approximately $-\lambda D$9 it/s for Reg-GAN (Xu et al., 2019).

In dual-control experiments, the method was illustrated on a logistic map and on cart-pole position control. For the logistic map with $s=\pm i$0, the first approximately $s=\pm i$1 steps were used for exploration and learning before the trajectory approached the target $s=\pm i$2; in the chaotic regime $s=\pm i$3, tracking remained feasible but required longer learning. In the cart-pole problem, the GP-based dual controller brought the trajectory within $s=\pm i$4 of the target within $s=\pm i$5 steps from $s=\pm i$6 (Alpcan, 2011).

Robust-control-oriented CLC has also produced explicit quantitative gains. In the distillation-column example for learning-enhanced robust controller synthesis, an a-priori sector bound $s=\pm i$7 yielded robust performance level $s=\pm i$8, whereas the learned sector bound $s=\pm i$9 led to robust performance level $s^2+\lambda s+1$0 with probability at least $s^2+\lambda s+1$1. In control-oriented qLPV identification, the reported BFR train/test scores improved from $s^2+\lambda s+1$2 to $s^2+\lambda s+1$3 on the HW benchmark, from $s^2+\lambda s+1$4 to $s^2+\lambda s+1$5 on MR, from $s^2+\lambda s+1$6 to $s^2+\lambda s+1$7 on Two-Tank, and from $s^2+\lambda s+1$8 to $s^2+\lambda s+1$9 on Silverbox when moving from an LTI baseline to a qLPV model (Fiedler et al., 2021, Mulagaleti et al., 2024).

Robotics examples emphasize physical interaction and safety. In safe whole-body loco-manipulation on a Unitree Go2 with a 6-DoF arm and wrist-mounted 6-DoF force/torque sensor, end-effector velocity tracking over about $L2$0 s of loco-manipulation achieved mean squared error $L2$1 for linear velocity components and $L2$2 for angular velocity components. The NN-augmented Kalman filter reported angular-velocity MAE/RMSE/STD of $L2$3 rad/s, $L2$4 rad/s, and $L2$5 rad/s on the $L2$6 axes, and linear-velocity $L2$7, $L2$8, and $L2$9 m/s, respectively (Schperberg et al., 2 Mar 2026).

Other CLC strands are more foundational or conceptual. Mission-aligned learning-informed control is presented through a robotic-care scenario involving meals, medication regimes, bathroom assistance, and entertainment, with RMPC responsible for safety and RL providing adaptation (Kungurtsev et al., 6 Jul 2025). Risk-sensitive RL under constraints formulates safety as the probability of hitting error states and applies a weighted performance-risk criterion to feed-tank control with stochastic inflows, outperforming an analytical chance-constrained solution when restrictive assumptions are relaxed (Geibel et al., 2011). Evolutionary deep RL with an elite buffer interleaves actor-critic learning with crossover, mutation, and elite-trajectory replay, reporting mean return $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$0 on HumanoidStandup-v2 and $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$1 on Swimmer-v2 (Esmaeeli et al., 2022).

5. Guarantees, stability, and safety

The strongest CLC results provide explicit stability or safety certificates, but the type of certificate varies sharply across formulations. In GAN stabilization, the Dirac GAN closed-loop matrix

$\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$2

is Hurwitz for $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$3, and the continuous-time Lyapunov equation $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$4 therefore admits a unique positive-definite solution for any $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$5, establishing exponential convergence. For normal GANs, the paper also states a local convergence theorem under mild regularity assumptions, small learning rate, and sufficiently large $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$6 (Xu et al., 2019).

In separation-based CLC for linear systems, the central theorem is optimality transfer: once the information state $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$7 is known, the separated policy designed offline on the nominal model is also optimal for the actual system. In the later LQR specialization, this proxy-to-real equivalence is sharpened by stage-specific penalties: when $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$8, one may choose $\Pi_t = p(X_t,\hat{X}_t \mid Y_{0:t}, U_{0:t-1}),$9 for all stages as $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$0; when $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$1, only $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$2 is universal and earlier $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$3 generally depend on the unknown dynamics (Malikopoulos, 2023, Kounatidis et al., 30 Sep 2025).

Robust and safety-critical CLC strands explicitly combine statistical and control-theoretic guarantees. In learning-enhanced robust controller synthesis, GP confidence bounds imply that the true uncertainty lies in the learned set with probability at least $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$4, and IQC-based synthesis then certifies robust stability and performance for all elements of that set. The main theorem states that, with probability at least $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$5, the synthesized controller stabilizes the true system and achieves the certified robust performance level $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$6 (Fiedler et al., 2021).

Hierarchical and robotic CLC frameworks typically express guarantees through feasibility and invariance. RMPC in mission-aligned autonomous systems provides worst-case constraint-satisfaction guarantees under learned dispersion sets $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$7, with safety holding “by construction” when the learned sets remain conservative with respect to data (Kungurtsev et al., 6 Jul 2025). In qLPV identification with RCI regularization, feasibility of the regularization QP certifies existence of an RCI set $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$8 and therefore of a robust constraint-satisfying controller for the uncertain learned model (Mulagaleti et al., 2024). In loco-manipulation, the Reference Governor modifies the reference pose and wrench using a precomputed admissible set and provides recursive feasibility and constraint-satisfaction guarantees under the assumed model used in the MOAS computation (Schperberg et al., 2 Mar 2026).

Not all CLC formulations provide such guarantees. The GP dual-control method states explicitly that stability is not guaranteed and that empirical behavior depends on the exploration weight and schedule (Alpcan, 2011). AntLer does provide almost-sure approximation guarantees for its sample-average approximation, showing that the approximate minimizer converges to the true optimum as the number of scenarios grows, but it does not itself establish Lyapunov safety guarantees (Capone et al., 2020).

6. Limitations, misconceptions, and open directions

A common misconception is that CLC denotes a single algorithmic family. The cited literature does not support that reading. The term is used for closed-loop stabilization of learning dynamics, dual control, robust synthesis with learned uncertainty, separation principles, hierarchical planning-MPC-RL, control-oriented identification, and even classical-quantum channel frameworks. This suggests that CLC is better read as a research nexus than as a fixed doctrine (Vedaie et al., 2023).

A second misconception is that “combining learning and control” automatically yields stronger guarantees. Several papers instead emphasize the opposite: dual control with GP regression offers no formal stability guarantee; HJB-PINN co-design does not provide formal convergence guarantees of PINNs to the viscosity solution; and the linear-system separation framework focuses on optimality rather than explicit closed-loop stability margins or regret bounds (Alpcan, 2011, Nishimura et al., 2024, Malikopoulos, 2023). Even where guarantees exist, they depend on restrictive assumptions such as smoothness, convexity of composite objectives, RKHS norm bounds, conservative uncertainty sets, or correct information-state recursion (Kounatidis et al., 30 Sep 2025, Fiedler et al., 2021).

Computational burden is another persistent limitation. GP-based methods incur $\beta_t (X_{t+1}-\hat{x}_{t+1})^2$9 covariance inversion cost as data grow (Alpcan, 2011). Mission-aligned frameworks must solve MPC and re-planning problems online (Kungurtsev et al., 6 Jul 2025). PINN-based HJB co-design requires large collocation sets, with $\beta_t$0 used at $\beta_t$1 (Nishimura et al., 2024). qLPV CLC differentiates through a convex QP inside training, and evolutionary DRL variants maintain a population of replay buffers and policies rather than a single learner (Mulagaleti et al., 2024, Esmaeeli et al., 2022).

Open directions are correspondingly diverse. The cited works point to Z-transform or discrete-time analyses for learning dynamics, multi-step MPC formulations with entropy or variance bonuses, safe exploration under RMPC, sparse or inducing-point GP variants for scalability, actor-critic or differentiable-argmin replacements when HJB minimizers lack closed forms, online adaptation of safety sets, and stronger convergence theory for hierarchical two-timescale learning (Xu et al., 2019, Alpcan, 2011, Kungurtsev et al., 6 Jul 2025, Nishimura et al., 2024, Mulagaleti et al., 2024). A plausible overarching implication is that future CLC research will continue to revolve around one central question: how much of the adaptability of learning can be retained once the structural demands of control—stability, constraint satisfaction, interpretability, and real-time feasibility—are made explicit.