Abstract: In this paper, we discuss an approximation strategy for solving the Linear Quadratic Tracking (LQT) problem that is both forward and local in time. We exploit the known form of the value function along with a time reversal transformation that addresses boundary condition consistency. We provide the results of an experimental investigation showing how the proposed solution performs relative to the optimal solution. Finally, we demonstrate that the proposed solution serves as a valid alternative to model predictive control (MPC) strategies, dramatically reducing computational burden.
Introduction
In optimal control, specifically the Linear Quadratic (LQ) framework, researchers face a fundamental "non-locality" challenge. To determine the optimal control strategy at any given moment, one technically requires knowledge of the system's state at the end of the time horizon. Mathematically, this requirement manifests in the Riccati differential equation, which must be solved backward from a final boundary condition.
For real-world applications involving long time horizons or high-dimensional state spaces, solving backward is often computationally prohibitive. The current industry standard, Model Predictive Control (MPC), attempts to circumvent this by solving optimization problems over a small "receding" window of the future. However, MPC remains computationally intensive and still requires some foresight into future reference signals. This paper proposes a "Forward Approximate Solution" that enables real-time, forward-only computation using only current information.
Methodology
The proposed strategy rests on two primary pillars: the Sign Flip Transformation and a Forward Tracking Approximation.
Sign Flip Transformation (SIFT)
The authors observe that the Riccati differential equation is autonomous. By applying a time-reversal transformation (defined as $\tau = T - s$), the equation can be reformulated to be solved forward in time. In the scalar case, the authors prove that this "flipped" equation is asymptotically stable. This stability ensures that the equation converges toward the optimal steady-state solution, turning a traditional mathematical hurdle into a computational advantage.
Forward Tracking Approximation
Standard Linear Quadratic Tracking (LQT) requires knowledge of the future reference signal $z(t)$ to solve for the tracking term ($\eta$). To achieve a local-in-time solution, the authors introduce an approximation for $\eta$ that utilizes only the current value of the reference signal. This allows the controller to update the trajectory entirely in a forward direction without "peeking" into future data points.
Experimental Setup
The algorithm was evaluated in a scalar LQT environment ($A=B=Q=1$) using three distinct synthetic reference signals designed to test various signal characteristics:
- $z_1(s)$: A continuous, periodic cosine wave that transitions to a constant value halfway through the horizon.
- $z_2(s)$: A non-periodic signal featuring 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 specifically chosen to evaluate how the forward approximation handles continuity, periodicity, and sudden jumps in the reference trajectory.
Key Insights and Results
Stability and Convergence
A critical insight of this work is the trade-off regarding the "initial gap." Because the forward approximation starts from a zero-condition rather than the true optimal condition inherited from the future, it is least accurate at the very beginning of the time interval. However, as time progresses, the SIFT-based solution quickly converges toward the optimal path.
Performance vs. Model Predictive Control
The experimental results highlight significant advantages over traditional MPC:
- Computational Efficiency: In a test case with a 250-second horizon, the Forward Approximate Solution required approximately 0.4 seconds to compute. In contrast, MPC required nearly 330 seconds to achieve a comparable error rate.
- Accuracy in Long Horizons: For long time horizons ($T=2000s$), the percentage error (PE) of the forward approximation was as low as 1.26% compared to the true optimal solution.
- Reliability: While MPC solutions often diverged (exceeding error values of $10^4$) during short horizons or under specific parameters, the Forward Approximate Solution remained stable across all tested reference signals.
Limitations and Future Directions
While promising, the current study focuses primarily on the scalar case ($n=1$). Although the underlying logic suggests the approach is applicable to multidimensional systems, formalization and experimental validation for complex, multi-variable controllers remain for future work. Additionally, the "initial gap" transient makes the method less suitable for extremely short-duration tasks where the initial error might dominate the total cost.
Future research will focus on extending the SIFT framework to multidimensional cases and formalizing the approach for discrete-time computational settings, which are standard in modern digital hardware.
Summary updated