Abstract
We propose a stochastic MPC scheme using an optimization over the initial state for the predicted trajectory. Considering linear discrete-time systems under unbounded additive stochastic disturbances subject to chance constraints, we use constraint tightening based on probabilistic reachable sets to design the MPC. The scheme avoids the infeasibility issues arising from unbounded disturbances by including the initial state as a decision variable. We show that the stabilizing control scheme can guarantee constraint satisfaction in closed loop, assuming unimodal disturbances. In addition to illustrating these guarantees, the numerical example indicates further advantages of optimizing over the initial state for the transient behavior.
Introduction
Stochastic Model Predictive Control (SMPC) is a powerful framework for controlling systems subject to random disturbances while ensuring that constraints are met with a specified probability. However, a significant challenge arises when dealing with unbounded noise, such as Gaussian disturbances. In these scenarios, a large noise spike can push the system state into a region where no feasible control sequence exists, causing standard SMPC algorithms to fail—a phenomenon known as the "Achilles' heel" of stochastic control.
The SMPC-ic Approach
To address the problem of infeasibility, we introduce SMPC with Initial State Optimization (SMPC-ic). The core innovation is treating the initial state of the predicted trajectory ($z_0$) as a decision variable rather than a fixed measurement ($x(k)$). By allowing the controller to "nudge" its starting point, we maintain feasibility even under extreme noise.
Technical Components
- State Decomposition: The system state is decomposed into a nominal plan (the "plan") and an error component (the "noise").
- Constraint Tightening: We employ Probabilistic Reachable Sets (PRS) to create safety buffers around the nominal trajectory, ensuring that the actual state remains within constraints despite stochastic perturbations.
- Initial State Optimization: The nominal initial state $z_0$ is constrained to lie on a line segment connecting the current measurement $x(k)$ and the previous prediction $z_1(k-1)$. This "line constraint" ensures recursive feasibility; if a solution existed at the previous time step, a solution is guaranteed to exist at the current step, regardless of the noise magnitude.
- Unimodal Assumption: By assuming noise follows a central convex unimodal distribution (a broad class including Gaussian distributions), we provide rigorous closed-loop safety guarantees.
Experimental Evaluation
The proposed scheme was evaluated using a DC-DC converter regulation benchmark under high-variance Gaussian noise. Performance was compared against a standard Linear Quadratic Regulator (LQR) and a traditional "fallback" or "backup" SMPC strategy (SMPC-bak) via 250,000 Monte Carlo simulations.
Key Results
- Safety and Reliability: As shown in the table above, LQR fails to handle noise, violating constraints nearly 95% of the time. In contrast, SMPC-ic maintains violations at approximately 10.8%, adhering to the user-defined chance constraints.
- Performance and Efficiency: SMPC-ic achieved a lower relative cost (109.76%) compared to the fallback strategy (112.95%), indicating more efficient control during transient phases.
- Computational Speed: SMPC-ic was roughly 39% faster on average than the fallback strategy (6.7ms vs 9.3ms) because it avoids the "solve-fail-retry" loop common in backup methods.
Visualizing System Behavior
The qualitative advantages of initial state optimization are evident in the system's phase trajectories. Unlike backup strategies that can cause "jittery" control by jumping between measurements and old predictions, SMPC-ic allows the controller to "slide" smoothly along the constraint boundaries.
As seen in the figure, the use of PRS (the green circles) provides a buffer that prevents the system from crossing the red constraint line at $x_1 = 2$. The optimization approach ensures that these transitions are smooth, eliminating the undesirable jumps often seen in traditional SMPC implementations.
Conclusion
By optimizing over the initial state, the proposed SMPC-ic scheme provides a robust solution to the problem of unbounded noise in stochastic control. It guarantees recursive feasibility and closed-loop constraint satisfaction while offering computational and performance benefits over traditional fallback strategies. Future work will explore extending these guarantees to nonlinear systems and reducing conservatism through dynamic constraint tightening.
Summary updated