Stochastic Model Predictive Control using Initial State Optimization

Henning Schlüter, Frank Allgöwer

TL;DR

This paper introduces a novel Stochastic Model Predictive Control (SMPC) scheme that treats the initial state of the predicted trajectory as an optimization variable rather than a fixed measurement. By allowing the controller to "nudge" its starting point along a specific line, the authors solve the chronic problem of infeasibility caused by unbounded noise (like Gaussian disturbances) while providing rigorous mathematical guarantees for stability and safety.

What Problem Does This Paper Address?

In standard Model Predictive Control (MPC), the controller looks at the current state of a system and plans a sequence of actions. However, in the real world, systems are hit by random disturbances (stochastic noise). Stochastic MPC tries to handle this by ensuring constraints are met with a certain probability (e.g., "the voltage stays below 5V at least 95% of the time").

The "Achilles' heel" of SMPC is unbounded noise. If your system is hit by a massive, unexpected spike (common in Gaussian distributions), the measured state might end up in a position where no valid control sequence exists—the math simply "breaks" (infeasibility). Existing solutions usually involve a "fallback" strategy: if the current measurement is infeasible, use the state you predicted you'd be in yesterday. This often leads to "jittery" control and conservative performance.

Methods

The authors propose SMPC with Initial State Optimization (SMPC-ic). Key technical components include:

  1. State Decomposition: They split the system state into a "nominal" part (the plan) and an "error" part (the noise).
  2. Constraint Tightening: They use Probabilistic Reachable Sets (PRS) to shrink the allowed boundaries of the nominal plan, leaving a "buffer" for the noise.
  3. Initial State Optimization: Instead of forcing the nominal plan to start exactly at the current measurement $x(k)$, they introduce a decision variable $\xi$. The nominal initial state $z_0$ is constrained to lie on a line connecting the actual measurement and the previous prediction.
  4. Unimodal Assumption: They assume the noise follows a "central convex unimodal" distribution. This is a broad class that includes the common Gaussian (Bell curve) distribution and allows them to prove that their safety buffers actually work in a closed loop.

Datasets

The paper does not use a large-scale external dataset. Instead, it evaluates the algorithm using a standard control engineering benchmark:

  • DC-DC Converter Regulation: A linear time-invariant system model used to simulate voltage regulation under high-variance Gaussian noise.

Key Insights and Ideas

  • The "Line" Constraint: By constraining the initial state to a line between the measurement and the previous prediction, the authors ensure recursive feasibility. If the problem was solvable yesterday, it is guaranteed to be solvable today, even if a massive noise spike occurs.
  • Smoothness over Fallbacks: Unlike "backup" strategies that flip-flop between the measurement and a prediction, this optimization approach allows for a smooth interpolation. This eliminates the "jitter" often seen when a controller suddenly ignores a measurement to maintain feasibility.
  • Computational Efficiency: Surprisingly, optimizing the initial state is often faster than fallback methods. Fallback methods often require trying to solve the problem with the measurement first, failing, and then re-solving with the backup. SMPC-ic solves a single optimization problem that covers both cases.

Key Results

The authors compared their SMPC-ic against a standard LQR (Linear Quadratic Regulator) and a SMPC-bak (the fallback/backup approach) via 250,000 Monte Carlo simulations:

  • Safety: While LQR violated constraints 94.47% of the time (as it ignores noise), SMPC-ic kept violations to ~10.8%, respecting the probabilistic "chance constraints" defined by the user.
  • Performance: SMPC-ic achieved a lower relative cost (109.76%) compared to the fallback strategy (112.95%), indicating more efficient control during transient phases.
  • 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.
  • Reduced Jitter: Qualitative plots show that SMPC-ic "slides" smoothly along constraints, whereas the fallback method exhibits undesirable jumps.

Limitations and Future Work

  • Linear Systems: The current proof and method are tailored for linear time-invariant systems. Extending this to complex nonlinear systems remains a challenge.
  • Conservatism: The use of fixed-size "tubes" (PRS) for constraint tightening can be conservative. The authors suggest that "growing tubes"—where the safety buffer changes over the prediction horizon—could improve performance.
  • Unimodal Requirement: While many distributions are unimodal, the mathematical guarantees rely strictly on this assumption; highly irregular or multi-modal noise distributions might break the safety proofs.

Summary updated

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.
BibTeX
@inproceedings{schluter2022stochastic,
  author = {Henning Schlüter and Frank Allgöwer},
  title = {{Stochastic Model Predictive Control using Initial State Optimization}},
  booktitle = {IFAC-PapersOnLine},
  year = {2022},
  doi = {10.1016/j.ifacol.2022.11.095},
}

Loading related content...

Figure