Adaptive control
Adaptive control is a control methodology in which the controller adjusts its own parameters in real time to compensate for uncertainties, changes, or disturbances in the system being controlled. Unlike fixed-gain controllers, which assume that the system's dynamics are known and constant, adaptive controllers learn from the discrepancy between expected and observed behavior, modifying their strategy to maintain performance.
The core idea is simple: if the system does not behave as the model predicts, the controller treats this mismatch as information and updates its internal representation of the system. This makes adaptive control essential for systems where the dynamics are unknown a priori, time-varying, or subject to environmental changes that cannot be anticipated. Aircraft at different altitudes, robotic systems handling unknown loads, and process control in chemical plants with varying feedstock compositions are all natural applications.
Two classical architectures
Model Reference Adaptive Control (MRAC): The designer specifies a reference model — a desired input-output behavior that the closed-loop system should emulate. The adaptive controller adjusts its parameters so that the actual system's output tracks the reference model's output. The adaptation law is typically derived from a stability criterion such as the Lyapunov method or the MIT rule. MRAC is elegant but fragile: if the reference model is too ambitious relative to the plant's actual capabilities, the adaptation can become unstable.
Self-Tuning Regulators (STR): The controller maintains an online estimate of the system's parameters using recursive identification (e.g., recursive least squares). At each step, the controller redesigns itself using the current parameter estimate as if it were the true model. This is an indirect adaptive scheme: identify, then control. The self-tuning approach is more flexible than MRAC but computationally heavier, and it can suffer from the 'identification/control conflict' — the controller's excitation may be insufficient for good identification, and good identification may require excitation that degrades control performance.
Stability and the fundamental trade-off
The deepest challenge in adaptive control is stability under adaptation. A controller that modifies itself is a nonlinear dynamical system even when the plant is linear. The interaction between the identification process and the control process can produce unexpected behaviors: parameter drift, burst phenomena (periods of good performance interrupted by sudden instability), and convergence to incorrect parameter values that happen to stabilize the loop temporarily.
This is why adaptive control theory has spent decades developing global stability proofs. The landmark result is that for linear systems with known structure but unknown parameters, adaptive controllers can be designed that guarantee stability under persistent excitation — a condition that the input signal must contain enough frequency content to excite all modes of the system. Without persistent excitation, the controller cannot learn what it needs to learn, and the adaptation becomes a form of open-loop estimation that may diverge.
Adaptive control in the broader systems landscape
Adaptive control sits at the intersection of control theory, system identification, and learning. It predates modern machine learning by decades but shares the same fundamental problem: how to act well in a system whose dynamics are not fully known. The difference is that adaptive control emphasizes online performance guarantees — stability, convergence, bounded error — whereas machine learning often emphasizes offline predictive accuracy and relies on separate safety systems for deployment.
The modern convergence of the two fields is visible in adaptive model predictive control, where a neural network learns the system dynamics online and a model predictive controller uses the learned model to plan. This hybrid approach inherits the strengths of both: the neural network captures complex, nonlinear, unknown dynamics; the MPC layer provides constraint satisfaction and stability guarantees. The constraint qualification of the MPC subproblem, the persistent excitation of the identification, and the stability of the overall loop are the three analytical pillars that must hold simultaneously — a systems design problem that requires tools from optimization, control, and learning theory together.