Dual control problem

The dual control problem is the fundamental tension in adaptive and predictive control systems between two competing objectives: regulation (keeping the system near a desired state) and identification (exciting the system sufficiently to learn its dynamics). A controller that only regulates may never collect the data needed to improve its model; a controller that only excites may drive the system into undesirable states. The problem is dual because the two objectives are coupled: better models enable better regulation, but better regulation may preclude model improvement.

The problem was first formalized by Alexander Fel'dbaum in the 1960s and remains unsolved in its general form. Optimal dual control requires solving a dynamic programming problem over the joint space of states and beliefs — a space whose dimensionality grows exponentially with time. For all but the simplest systems, the optimal dual control policy is computationally intractable.

In practice, engineers use suboptimal approximations: certainty equivalence (ignoring uncertainty and acting as if the current model is exact), probing (adding small perturbations to improve identification), and separate excitation experiments (collecting data offline and then switching to regulation). Each approximation trades off performance for tractability, and none is universally adequate.

The dual control problem is not merely a technical challenge in control theory. It is a formal statement of the exploration-exploitation tradeoff that appears in machine learning, reinforcement learning, and active inference. Any system that must simultaneously act on its current knowledge and improve that knowledge faces the dual control problem. The framing is general enough to apply to scientific experimentation, institutional reform, and personal decision-making: how much should you disrupt the current state to learn whether a better state is possible?