What Are Neurocontrollers?

Neurocontrollers are control systems that use artificial neural networks as their primary computational element to generate control actions for dynamic plants, replacing or supplementing the classical analytical controllers derived from linear control theory. By learning input-output mappings from data rather than requiring an explicit mathematical model of the controlled system, neurocontrollers can handle nonlinear, time-varying, and poorly characterized dynamics that challenge traditional proportional-integral-derivative and model-predictive controllers. The field sits at the intersection of control engineering, machine learning, and optimization theory, drawing techniques from adaptive control, reinforcement learning, and neural network training.

Neurocontrollers emerged as a practical research area in the late 1980s and early 1990s as backpropagation training became established, enabling neural networks to learn stable control policies through iterative interaction with simulated or real plant models. They range from feedforward networks that compute static input-to-output mappings to recurrent architectures that maintain internal state for controlling systems with memory.

Architecture and Learning

The most common neurocontroller architectures pair a neural network controller with a system identifier, a second neural network trained to approximate the plant dynamics. The identifier provides a differentiable model of the plant that allows the controller's weights to be updated by backpropagating through the combined controller-identifier system. In direct inverse control, the network learns the inverse input-output map of the plant and generates the control input needed to produce a desired output. Adaptive critic designs split the task between an action network, which generates control commands, and a critic network, which evaluates long-run cost, paralleling the actor-critic formulation in reinforcement learning. IEEE research on robust adaptive critic-based neurocontrollers demonstrates how this two-network structure handles plant uncertainties while providing stability guarantees through Lyapunov-based analysis.

Stability and Performance

Stability analysis for neurocontrollers is more involved than for linear controllers because neural networks introduce nonlinear, data-dependent dynamics into the closed-loop system. Lyapunov stability methods have been adapted to provide conditions under which a neurocontroller will drive tracking error to zero despite bounded disturbances and parameter uncertainty. Adaptive neural network schemes that update weights online while the controller operates can achieve uniform ultimate boundedness of closed-loop errors, as shown in foundational work on adaptive neural network control of nonlinear systems. Radial basis function networks have been particularly popular in stability-guaranteed designs because their localized basis functions limit how much a new training sample disturbs already-learned regions of the input space, reducing the risk of catastrophic forgetting during online adaptation.

Hardware Implementation

Deploying neurocontrollers on embedded platforms requires mapping the trained network onto hardware with tight real-time, power, and memory constraints. Microcontrollers and digital signal processors have been used for lower-complexity networks, while field-programmable gate arrays enable the parallel arithmetic required for larger networks at the update rates demanded by fast mechanical or power electronic plants. Fixed-point arithmetic and weight quantization reduce memory bandwidth without unacceptable accuracy loss, and hardware-software co-design tools automate the mapping from floating-point training to efficient integer inference. Research on adaptive neurocontrollers for drive systems surveys the practical implementation strategies used to deploy neural network controllers on motor drive hardware, including training schedules and fault-tolerance considerations.

Applications

Neurocontrollers have applications in a wide range of disciplines, including:

• Industrial robotics, for controlling manipulators with uncertain payload and kinematic parameters
• Electric motor drives and power converters requiring fast, nonlinear torque and voltage regulation
• Autonomous vehicle path tracking and throttle-brake control under varying road conditions
• Chemical and process engineering, controlling reactors with complex nonlinear dynamics
• Aerospace flight control for aircraft with actuator faults or highly flexible structures