Combining Prediction and Reaction in Control Systems

43 concepts33 papers1 proposal

Imagine trying to steer a massive ship through a narrow, winding channel. If you only look directly over the bow, you can only react to obstacles after you encounter them. Because a ship takes time to turn, this purely reactive approach leads to overcorrection, dangerous oscillations, and eventual disaster. Conversely, if you steer blindly by a pre-computed map without ever looking out the window, an unexpected current will quietly push you off course.

The optimal solution is intuitive: look ahead to plan your main turns, but constantly check your position to make minor corrections.

In control theory, this is the essence of hybrid feedforward-feedback control. Pure feedback is reactive and robust, but it suffers from latency. Pure feedforward is proactive and instantaneous, but it is blind to unexpected disturbances. By combining the two, systems achieve a level of precision and stability that neither paradigm can provide alone. As David A. Robinson demonstrates in Information Processing to Create Eye Movements, even biological neural networks naturally blur the line between feedforward and feedback loops to manage complex, messy physical realities.

The Strategic Division of Labor

Why does this hybrid architecture work so well? It begins with a clean division of labor: the feedforward path handles the "heavy lifting"—the predictable physics, known trajectories, or large-scale structural shifts—while the feedback loop cleans up the residual errors and unmodeled noise.

Consider a high-speed wind tunnel where pressure must be precisely maintained during supersonic tests. In Feedforward-feedback compound control method and application for flow field in 2 meter high-speed free jet wind tunnel, Weijie Liu et al. use a predictive feedforward model based on physical flow conservation laws to generate the bulk of the control command. Because this model handles the massive, predictable aerodynamic shifts, a reactive variable-gain PI feedback loop only needs to manage minor deviations, keeping tracking errors under 0.3%.

[Desired State] ---> [ Feedforward Model (Predicts Bulk Command) ] ---\
               \                                                       (+) ---> [ Plant ]
                \--> [ Feedback Loop (Corrects Residual Error) ] -----/

This division of labor is equally vital when dealing with highly nonlinear, "mushy" actuators like those found in soft robotics. Hongbo Zhang et al. demonstrate in Control of Pneumatic Artificial Muscles with SNN-based Cerebellar-like Model that a spiking neural network can act as a feedforward component to predict the complex pressure requirements of pneumatic muscles. Meanwhile, a traditional proportional-derivative (PD) feedback loop catches immediate environmental disturbances.

By offloading the primary effort to a feedforward command, the system minimizes how aggressively the feedback loop must intervene. In fact, Pilhwa Lee et al. show in On the Existence and Computation of Minimum Attention Optimal Control Laws that an optimal "minimum attention" strategy relies almost entirely on open-loop, time-varying feedforward terms for the bulk of an initial movement, only ramping up precise feedback gains as the system nears its final target.

Defeating Latency and Dead Time

In physical systems, information takes time to travel. Sensory processing, mechanical inertia, and computational delays introduce "dead time." If a controller relies solely on feedback, it is always acting on old data.

Feedforward control defeats latency by acting proactively. In the foundational paper The Discrete Optimal Finite Preview Control Problem, Masayoshi Tomizuka formalized "preview control." By utilizing a short window of future information—akin to a driver spotting a pothole a few meters ahead—the optimal control law combines a standard feedback term with a weighted sum of future preview data. This allows the system to prepare for disturbances before they physically hit.

When future information cannot be directly observed, the system can instead simulate it using an internal forward dynamics model. In Evidence for a Forward Dynamics Model in Human Adaptive Motor Control, Nikhil Bhushan et al. show that the human central nervous system uses an internal simulator to predict the sensory consequences of motor commands. Rather than waiting for delayed sensory feedback from the limbs, the brain makes real-time corrections based on these predicted states.

This predictive-reactive synergy is highly effective across engineering domains:

Co-Adaptation and Learning

The true power of hybrid control emerges when the feedforward and feedback components are not just running in parallel, but actively communicating and adapting to one another.

In biological systems, feedback acts as a teacher. Paul Dean et al. model this in A Neural Net Model for Adaptive Control of Saccadic Accuracy by Primate Cerebellum and Brainstem. They show how a low-precision feedback controller in the brainstem provides immediate baseline responses, while an adaptive feedforward neural network in the cerebellum observes the resulting errors to calibrate its own gains. Over time, the feedforward network learns to handle the complex nonlinearities of eye muscles perfectly. A similar mechanism is explored by Eimei Oyama et al. in Coordinate Transformation Learning of Hand Position Feedback Controller by Using Change of Position Error Norm, where a feedback loop refines coordinate transformations based on real-time error changes to improve feedforward tracking.

In industrial engineering, neural networks are frequently deployed alongside classical controllers to handle high-speed transients. In Control Method for Improving Voltage Response and Ripple of Isolated DC-DC Converter, Ming Li et al. integrate a neural network as a high-speed auxiliary compensator alongside a traditional proportional-integral (PI) controller. During sudden load changes, the neural network instantly calculates duty ratio compensations, slashing voltage recovery times from seconds to milliseconds, while the PI controller ensures long-term steady-state stability.

This concept of combining rapid feedforward mapping with recurrent error-feedback is generalized by Ke Liu et al. in An Integrated Architecture of Adaptive Neural Network Control for Dynamic Systems, demonstrating a smoother and more accurate response than either method could achieve in isolation.

[Error Signal] ---> [ Feedback Controller ] ---> [ Learning Mechanism ]
                                                          |
                                                          v
[Input Signal] ---> [ Adaptive Feedforward Network (Learns Dynamics) ]

Even in advanced algorithmic frameworks like Model Predictive Control (MPC) and Reinforcement Learning (RL), cooperation is key. Yunyi Jia et al. introduce a cooperative framework in Actor-Critic Cooperative Compensation to Model Predictive Control for Off-Road Autonomous Vehicles Under Unknown Dynamics. Instead of letting the predictive MPC and the learning-based RL compensator fight for control, the authors include the compensator's output directly within the MPC's state vector. This allows the primary predictive controller to anticipate the learning agent's adjustments, preventing conflicting signals and ensuring smooth navigation over unpredictable terrain.

As Onno Eberhard et al. note in A Pontryagin Perspective on Reinforcement Learning, while open-loop planning offers massive computational advantages, an integrated feedback layer remains essential to guarantee robustness against real-world model inaccuracies.

Finally, this synergy can even be hardcoded directly into hardware. In VLSI Model of Primate Visual Smooth Pursuit, Ralph Etienne-Cummings et al. implement a silicon chip that combines negative feedback to minimize immediate tracking errors with positive feedback that stores velocity as a predictive feedforward signal, allowing the system to maintain stable tracking even when the target is briefly obscured.

By pairing the foresight of feedforward models with the oversight of feedback loops, hybrid control architectures deliver the best of both worlds: immediate, proactive speed and robust, reactive accuracy.

Key Foundational Work

The Continuous Optimal Finite Preview Control Problem (1976) — contributed 2 concept elements to this direction.
The Discrete Optimal Finite Preview Control Problem (1975) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hybrid_feedforward_feedback_control_072828a4,
  title = {hybrid feedforward-feedback control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-feedforward-feedback-control/rdr_072828a4692d4e0a848a087cb53c0f70},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_hybrid_feedforward_feedback_control_072828a4_v1,
  title = {hybrid feedforward-feedback control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-feedforward-feedback-control/rdr_072828a4692d4e0a848a087cb53c0f70/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake