Forward Approximate Solution for Linear Quadratic Tracking

Alessandro Betti, Michele Casoni, Marco Gori

TL;DR

This paper introduces a "Forward Approximate Solution" for Linear Quadratic Tracking (LQT) that eliminates the need to solve complex equations backward in time. By applying a Sign Flip Transformation (SIFT), the authors create a local-in-time, forward-only computational scheme that is orders of magnitude faster than traditional Model Predictive Control (MPC) while maintaining high accuracy for long-horizon tasks.

What Problem Does This Paper Address?

In optimal control, particularly the Linear Quadratic (LQ) problem, we face a "non-locality" challenge. To find the best control strategy today, you technically need to know the state of the system at the very end of the time horizon. Mathematically, this manifests as the Riccati equation, which must be solved backward from a final boundary condition.

For real-world applications with long time horizons or high-dimensional variables, solving backward is often computationally impossible. The industry standard is Model Predictive Control (MPC), which approximates the solution by looking at a small "receding" window of the future. However, MPC is computationally expensive because it requires solving an optimization problem at every single time step and still requires some knowledge of future reference signals. The authors seek a way to solve these problems forward, in real-time, using only current information.

Methods

The authors propose a strategy based on two main pillars:

  1. Sign Flip Transformation (SIFT): They observe that the Riccati differential equation is autonomous. By performing a time-reversal transformation ($\tau = T - s$), the equation can be reformulated to be solved forward in time. In the scalar case, they prove this "flipped" equation is asymptotically stable, meaning it converges to the optimal steady-state solution.
  2. Forward Tracking Approximation: For the tracking problem (LQT), where the system tries to follow a reference signal $z(t)$, the standard math requires knowing the future of that signal. The authors introduce an approximation for the tracking term ($\eta$) that uses the current value of the reference signal. This allows the controller to update the trajectory entirely forward in time without "peeking" into the future.

The resulting algorithm is "local in time," meaning it only needs the current state and current reference signal to compute the next control action.

Datasets

The paper uses synthetic reference signals to evaluate the algorithm's performance in a scalar LQT environment ($A=B=Q=1$):

  • $z_1(s)$: A continuous, periodic cosine wave that becomes constant halfway through.
  • $z_2(s)$: A non-periodic signal with a sharp discontinuity, followed by a sigmoidal function.
  • $z_3(s)$: A non-periodic signal with a discontinuity, followed by a constant value.

These signals were chosen to test how the approximation handles continuity, periodicity, and sudden jumps.

Key Insights and Ideas

  • Stability via Time-Reversal: The most elegant insight is that while the standard Riccati equation is unstable when solved forward, its time-reversed counterpart (SIFT) is stable. This turns a mathematical hurdle into a computational advantage.
  • The "Initial Gap" Trade-off: The forward approximation is least accurate at the very beginning of the time interval (the "initial gap") because it starts from a zero-condition rather than the "true" optimal condition inherited from the future. However, as time progresses, the solution quickly converges toward the optimal path.
  • MPC vs. Forward Approach: The authors show that if MPC is used in a strictly "online" manner (without future info), it often diverges or fails. Their forward approach, by contrast, remains stable and reliable even without future knowledge.

Key Results

  • Computational Speedup: The results are dramatic. In a test case with a 250-second horizon, the Forward approach took approximately 0.4 seconds to compute, while MPC required nearly 330 seconds to achieve a comparable error rate.
  • Accuracy: For long time horizons ($T=2000s$), the percentage error (PE) of the forward approximation dropped to as low as 1.26% compared to the true optimal solution.
  • Reliability: Table 2 shows that for short horizons or specific parameters, MPC solutions often diverged (marked as "$*$" for values $> 10^4$), whereas the Forward Approximate Solution remained stable across all tested reference signals.

Limitations and Future Work

  • Scalar Focus: The current paper focuses primarily on the scalar case ($n=1$). While the authors suggest the logic holds for multidimensional systems, they haven't yet provided the full formalization or experimental proof for complex, multi-variable controllers.
  • The Starting Transient: Because the method is an approximation that ignores the final boundary condition, there is always an initial period of higher error. This might make it unsuitable for extremely short-duration tasks where the "initial gap" dominates the total cost.
  • Future Directions: The authors plan to extend this to multidimensional cases and formalize the approach for discrete-time computational settings, which are more common in digital hardware.

Summary updated

Abstract

In this paper, we discuss an approximation strategy for solving the Linear Quadratic Tracking that is both forward and local in time. We exploit the known form of the value function along with a time reversal transformation that nicely addresses the boundary condition consistency. We provide the results of an experimental investigation with the aim of showing how the proposed solution performs with respect to the optimal solution. Finally, we also show that the proposed solution turns out to be a valid alternative to model predictive control strategies, whose computational burden is dramatically reduced.
BibTeX
@inproceedings{betti2023forward,
  author = {Alessandro Betti and Michele Casoni and Marco Gori},
  title = {{Forward Approximate Solution for Linear Quadratic Tracking}},
  booktitle = {WIRN},
  year = {2023},
  doi = {10.1007/978-981-96-0994-9\_6},
}
Figure