The Control Revolution: Beyond Derivatives

What if the greatest obstacle to precision engineering isn’t the complexity of the machine, but the math we use to describe it? For decades, the "brain" of advanced machinery—Nonlinear Model Predictive Control (NMPC)—has relied almost exclusively on derivatives. These solvers require smooth, predictable curves to function.

However, the real world is often jagged, non-differentiable, and hidden inside "black-box" models that refuse to surrender their underlying equations.

A New Paradigm: Ditching Derivatives

A new framework is shifting this paradigm by ditching derivatives entirely in favor of a sophisticated search pattern. Researchers have successfully integrated the Mesh Adaptive Direct Search (MADS) algorithm into a control framework.

This discovery matters because it paves the way for autonomous systems—from drones to chemical plants—to make split-second decisions using high-fidelity simulations that were previously too "clunky" or complex for traditional controllers to process.

Key Components of the MADS-NMPC Framework

The Core Innovation

The framework replaces traditional derivative-based solvers with the Mesh Adaptive Direct Search (MADS) algorithm, enabling control of systems described by non-smooth or "black-box" models.

The Scalability Breakthrough

Unlike previous methods, the authors noted a critical advantage: "the number of points evaluated around the current iterate grows linearly with the input dimension, instead of combinatorially." This is achieved through an ingenious state-augmentation scheme that computes costs and constraints in a single pass.

A Practical Demonstration: Rocket Control

The Challenge

The team demonstrated the framework's power by tackling a robust rocket throttle control model. The goal was deceptively simple: hit a target apogee of 10,000 feet (3048 meters).

The Constraints

The system had to grapple with significant uncertainties while adhering to strict rules:

Uncertain parameters like the drag coefficient and specific impulse.
A strict terminal velocity constraint of v ≤ 150 m/s.
Using a shared thrust profile across multiple, independently timed trajectories.

The Results

The application of the MADS-based solver yielded stark and promising outcomes:

Precision: Reduced cumulative velocity violations from runaway levels to near-zero states.
Optimality: The solver naturally produced a "bang-singular-bang" thrust profile—a signature of physical optimality confirming the algorithm is mastering physics, not just crunching numbers.
Capability: Successfully managed the complex problem of independent final times for multiple trajectories.

Current Limitations & Future Challenges

Remaining Friction

The path to universal application is not without its own obstacles. The current implementation faces two primary hurdles:

Sensitivity to Initial Conditions: The reliance on a "single-shooting" method means the system is highly sensitive to its starting point; a tiny initial nudge can send a trajectory wildly off course.
Implementation & Scale: The Julia-based implementation in DirectSearch.jl faces minor hurdles in passing variables between calculation phases. Furthermore, while linear complexity growth is a massive upgrade, high-dimensional problems over long time horizons remain a formidable challenge.

Source: Towards a Framework for Nonlinear Predictive Control using Derivative-Free Optimization by Ian McInerney, Lucian Nita, Yuanbo Nie, Alberto Oliveri, and Eric C. Kerrigan. (arXiv:2106.05025v1)