Every human-machine system is a feedback system, whether its designers acknowledge it or not. A pilot holding altitude, an operator regulating coolant flow, a driver tracking a lane: each closes a loop through perception, cognition, and neuromuscular action. The human is not an external user of the system. The human is a controller embedded inside it, with finite bandwidth, inherent time delay, and noise that no amount of training fully eliminates.
Interface design is often treated as a matter of ergonomics and aesthetics, such as legible fonts, sensible layouts, and intuitive controls. Those matter. But they leave the central engineering question unanswered. Does the closed loop formed by operator, interface, and plant have adequate stability margins, acceptable tracking performance, and robustness to the disturbances it will actually encounter? Heuristics alone cannot answer that question. It requires a model of the operator in the same mathematical language as the rest of the loop.
Control theory supplies that language. Over seven decades, researchers from Tustin and McRuer to Kleinman and Baron have developed quantitative descriptions of human control behaviour that fit directly into frequency-domain and state-space analysis. Once the operator is modelled this way, interface design becomes a tractable optimisation problem. What follows examines three pillars of that approach. The first is modelling the human controller. The second is designing displays that make the system observable at reasonable cost. The third is allocating control functions between human and automation according to their dynamic characteristics.
Modelling the Human as a Controller
The foundational result came from studies of manual tracking in the 1950s and 1960s. Duane McRuer and colleagues found that skilled operators performing compensatory tracking adapt their own dynamics to the controlled element. Whatever the plant, the combined open-loop dynamics of human and plant near crossover take approximately the same form: an integrator, a crossover gain, and an effective time delay. This is the crossover model, and it remains one of the most robust empirical findings in engineering psychology.
The model's power lies in what it predicts about workload. If the plant is a pure gain, the operator behaves as an integrator, which is a comfortable, lagging response. If the plant is itself an integrator, as with rate-commanded vehicles, the operator need only supply gain. If the plant is a double integrator, typical of acceleration-commanded systems, the operator must generate lead by effectively differentiating the error signal. Humans can do this, but at a cost. The crossover frequency falls, effective delay grows, and subjective workload rises sharply.
The crossover model is quasi-linear. It captures the portion of operator output that is linearly correlated with the input. Everything else goes into a remnant, which is injected noise arising from perceptual imprecision, motor variability, and shifting strategy. Treating the human as a describing function plus remnant lets the standard tools apply: Bode analysis, phase margin computation, and closed-loop bandwidth estimation. It also clarifies a classic failure mode. Pilot-induced oscillation often emerges when added lag erodes phase margin at the frequencies where a high-gain operator closes the loop.
The state-space counterpart is the Optimal Control Model developed by Kleinman, Baron, and Levison around 1970. It assumes that a well-trained operator behaves approximately as an optimal controller subject to human limitations: perceptual delay, observation noise on each displayed variable, motor noise, and neuromuscular lag. The model contains a Kalman filter that estimates state from noisy observations, a predictor that compensates for delay, and a linear-quadratic regulator that generates control. This structure is not a claim about neural architecture. It is a claim that skilled behaviour converges toward what an optimal, constrained estimator-controller would do.
The two models are complementary. The crossover model is compact and well suited to early frequency-domain screening of handling qualities. The Optimal Control Model scales to multivariable tasks and shows explicitly how display quality, attention, and noise propagate to performance. Both models change how the operator is treated. Instead of an unpredictable element to be accommodated, the operator becomes a dynamic subsystem to be designed around, with known parameters, known limits, and known ways of failing.
TakeawayThe human operator is not the unpredictable part of the loop. It is an adaptive controller with a well-characterised bandwidth, delay, and noise floor, and those limits belong in the design specification.
Displays as Instruments of Observability
In classical control theory, a system is observable if its internal state can be reconstructed from its outputs. The rank test is binary: the state is either recoverable or it is not. For a human operator, this binary notion is not enough. A display showing only the position of a double integrator is technically observable, because velocity can be inferred from how quickly position changes. But humans estimate rates from visual motion with substantial noise and delay. Mathematical observability does not guarantee practical observability.
A more useful framing looks at the conditioning of the estimation problem. In the Optimal Control Model, each displayed variable carries observation noise scaled to its magnitude. The operator's internal filter must reconstruct any undisplayed states by differentiating or integrating what it can see. When critical states must be inferred indirectly, estimation error grows and the filter leans more heavily on its internal model. The consequences are measurable: tracking performance degrades and the effort required to sustain it rises. A well-designed display reduces this burden by presenting states the operator would otherwise have to derive.
Two classic techniques illustrate the trade-offs. Quickening, introduced by Birmingham and Taylor in the 1950s, displays a weighted sum of the error and its derivatives. The operator sees a signal that responds to control input like a simple integrator, so a high-order plant becomes easy to fly. The flight director is a direct descendant of this idea. The cost is transparency. The operator sees a synthetic command rather than the true state, and loses the ability to notice when the underlying dynamics have changed.
Predictor displays offer an alternative. They show the actual current state alongside a model-based projection of where the system will be several seconds ahead, given the current control input. This supplies the lead information the operator would otherwise have to generate internally. It also preserves the true state for situational awareness and failure detection. Submarine depth control and spacecraft rendezvous have both benefited from this approach. The design question becomes one of prediction horizon and model fidelity. A horizon that is too short gives insufficient lead, while one that is too long lets model error dominate.
Multivariable displays add another dimension. Senders' work on visual sampling showed that operators sample instruments at rates tied to signal bandwidth, broadly consistent with the Nyquist criterion. In the Optimal Control Model, attention is a finite resource, and observation noise on each channel scales inversely with the share of attention it receives. Display design then becomes a resource allocation problem. Integrating related variables into a configural display, or presenting derived quantities directly, reduces the number of channels competing for attention and lowers total estimation error for the same effort.
TakeawayA display is not a window onto the system. It is the measurement matrix of the operator's estimator, and designing it well means choosing which states the human should observe directly and which they should never have to derive.
Allocating Control by Dynamic Characteristics
The question of what humans should do and what machines should do is as old as automation itself. Fitts' 1951 list compared human and machine capabilities in qualitative terms: machines excel at speed and repetition, humans at judgement and improvisation. Later frameworks added granularity across stages of information processing, notably the levels of automation proposed by Sheridan and Verplank and extended by Parasuraman, Sheridan, and Wickens. Control theory contributes something these frameworks lack. It offers a quantitative basis for allocation grounded in the dynamics of the loops being closed.
The central principle is time-scale separation. A human operator's crossover frequency is bounded, typically to a few radians per second for well-practised tasks, and constrained by effective delays of a few hundred milliseconds. Any loop that must be closed above that bandwidth is a poor candidate for manual control. This is why modern aircraft use stability augmentation for fast modes such as the short period and Dutch roll, leaving the pilot to manage slower attitude and flight-path loops. The automation shapes the inner-loop dynamics so that the outer loop the human controls falls within the range where the crossover model predicts good performance.
Bandwidth is not the only dimension. Loops that are slow but involve ambiguous goals, unmodelled disturbances, or rare contingencies often benefit from human supervision, because the operator's adaptive capacity is exactly what the crossover model documents. The difficulty lies in the middle ground. These are loops fast enough to demand continuous attention but slow enough that either automation or human could plausibly close them. Here, allocation should be driven by the consequences of failure, the predictability of the disturbance environment, and the cost of transitions between modes.
Transitions are where allocation designs most often fail. Lisanne Bainbridge's 1983 paper on the ironies of automation identified the core problem. Automation tends to return control to the human precisely when conditions exceed its design envelope, which is when the human is least prepared to take over. In estimation terms, an operator who has been monitoring rather than controlling holds a degraded state estimate, because their internal filter has drifted. On handover, they must re-establish an accurate picture of the system while also generating control, under conditions that are abnormal by definition.
This suggests that allocation should be evaluated as a hybrid system rather than a static assignment, with discrete modes and continuous dynamics within each. Engineers can analyse mode transitions explicitly. They can ask what state the system is likely to be in at handover, how long the operator needs to rebuild an adequate estimate, and whether the plant's dynamics allow that time. Designs that keep the human partially engaged, through shared control or graceful degradation of authority, keep the operator's estimate current at the cost of some steady-state workload. That trade is often worth making.
TakeawayFunction allocation is not a decision about who is better at a task. It is the design of a hybrid dynamic system, and its most critical behaviour occurs at the boundaries between modes.
Viewed through control theory, human-machine interface design stops being a matter of preference and becomes a problem of closed-loop synthesis. The operator is a controller with characterisable dynamics. The display is a measurement system feeding that controller's estimator. The allocation of functions defines the architecture of a hybrid system whose modes and transitions can be analysed like any other.
This framing does not replace human factors research, cognitive engineering, or empirical testing. It complements them by providing a common language. In that language, handling qualities, display formats, and automation logic can be evaluated against the same performance criteria as the rest of the system. Trade studies become quantitative, and failure modes become predictable.
The methodology follows a consistent sequence. Model the operator. Shape the plant so the human's task falls within their bandwidth. Design displays that minimise the estimation burden. Then analyse every mode transition as rigorously as steady-state operation. The human remains the most adaptive element in the loop, and good design gives that adaptability room to work.