Data-Driven Tube-Based Stochastic Predictive Control

Sebastian Kerz, Johannes Teutsch, Tim Brüdigam, Marion Leibold, Dirk WollherrIEEE Open Journal of Control Systems 2021 · DOI · OpenReview

TL;DR

This paper introduces a novel data-driven control strategy for linear systems facing both unknown dynamics and unpredictable disturbances. By combining ideas from "tube-based" control and stochastic Model Predictive Control (MPC), the authors develop a method that can guarantee system stability and constraint satisfaction using only observed data, even with noisy measurements. This approach offers a computationally efficient way to achieve robust and probabilistic control for complex systems.

What Problem Does This Paper Address?

Traditional control methods often rely on accurate mathematical models of the system being controlled. However, in many real-world applications, these models are either unknown or too complex to derive. While data-driven control methods, particularly those based on Willems' fundamental lemma, have emerged as a powerful alternative for learning system behavior directly from data, they often focus on robustness against measurement noise. A significant challenge remains in effectively incorporating probabilistic information about disturbances to achieve performance-oriented control, akin to stochastic MPC, while still handling unknown dynamics and noisy measurements. This paper aims to bridge this gap by proposing a data-driven approach that leverages probabilistic disturbance information for performance optimization, while ensuring recursive feasibility and stability.

Methods

The core of the proposed method lies in a novel data-driven formulation of "tube-based" stochastic Model Predictive Control (MPC). The key ideas are:

  • Data-Driven System Representation: The authors utilize Willems' fundamental lemma to represent the system's dynamics using only observed input, output, and disturbance trajectories. This avoids the need for explicit system identification.
  • State Decomposition: The system's state is split into a deterministic "nominal" state (influenced by inputs) and a stochastic "error" state (affected by disturbances). This decomposition is crucial for making the optimization problem tractable.
  • Tube-Based Control: Inspired by tube-based MPC, the approach uses the deterministic nominal state for optimization. Constraints on the original state are then translated into tightened constraints on this nominal state.
  • Probabilistic and Robust Constraint Tightening: To satisfy chance constraints (i.e., constraints that must hold with a certain probability), the nominal state constraints are probabilistically tightened with respect to the additive disturbance. Additionally, these constraints are further tightened robustly to account for measurement noise. This ensures that original state constraints are met with a pre-specified probability level.
  • Receding Horizon Optimization: The control problem is solved in a receding horizon fashion, meaning the controller re-optimizes at each time step based on the latest measurements.
  • Data-Driven Formulation: All necessary components, including constraint sets and feedback gains, are computed directly from the available data.

Datasets

The paper uses simulated trajectory data for training and evaluation. Specifically:

  • Assumption 1: Requires a persistently exciting input-output-disturbance trajectory.
  • Assumption 2: Requires a set of available disturbance sequences and knowledge of a bound on measurement noise.
  • Simulation Example: A double-mass-spring-damper system is simulated to generate data for demonstrating the algorithm's performance. This simulation involves both additive disturbances and measurement noise.

Key Insights and Ideas

  • Bridging Data-Driven and Stochastic Control: The paper successfully combines the power of data-driven control (learning from data) with the performance benefits of stochastic control (handling probabilistic uncertainties).
  • Deterministic Reformulation of Stochastic Problems: By decomposing the state and using tube-based ideas, the authors transform a complex, intractable stochastic optimal control problem into a computationally feasible deterministic one.
  • Guaranteed Performance with Noisy Data: The proposed method provides theoretical guarantees for constraint satisfaction and system stability, even when measurements are corrupted by noise and the system dynamics are unknown.
  • Lightweight and Efficient: The resulting data-driven MPC problem is described as "lightweight," suggesting it can be solved efficiently in real-time applications.

Key Results

  • The proposed data-driven stochastic MPC scheme is shown to be recursively feasible, meaning if a solution exists at one time step, a solution will also exist at the next.
  • The closed-loop system is proven to be input-to-state stable (ISS) with respect to both additive disturbances and measurement noise.
  • A simulation example using a double-mass-spring-damper system demonstrates the effectiveness of the approach. The controller successfully stabilized the system while respecting input and state constraints, even under noisy conditions.
  • The paper shows that the proposed controller can handle inexact offline data, with performance degradation being relatively small in their tested scenarios.

Limitations and Future Work

  • Inexact Data Assumption: The paper acknowledges that the assumption of exact offline data (Assumption 1) is a simplification. While they investigate the effect of inexact data, the theoretical guarantees might not hold directly. Future work could focus on developing methods to handle inexact data more rigorously.
  • Prediction Error Bounds: The paper mentions that bounding prediction errors in the presence of noisy data is an open challenge.
  • Regularization: For inexact data scenarios, the authors suggest adding regularization terms to the cost function to penalize large decision variables, which can help mitigate noise amplification.
  • Generalization to Nonlinear Systems: The current work focuses on linear time-invariant (LTI) systems. Extending the approach to nonlinear systems would be a valuable future direction.
  • Computational Complexity: While described as "lightweight," the computational cost of solving the OCP might still be a limiting factor for very fast systems or long prediction horizons. Further optimization of the solver or algorithm could be beneficial.

Summary updated

Abstract

A powerful result from behavioral systems theory known as the fundamental lemma allows for predictive control akin to Model Predictive Control (MPC) for linear time-invariant (LTI) systems with unknown dynamics purely from data. While most data-driven predictive control literature focuses on robustness with respect to measurement noise, only a few works consider exploiting probabilistic information of disturbances for performance-oriented control as in stochastic MPC. This work proposes a novel data-driven stochastic predictive control scheme for chance-constrained LTI systems subject to measurement noise and additive stochastic disturbances. In order to render the otherwise stochastic and intractable optimal control problem deterministic, our approach leverages ideas from tube-based MPC by decomposing the state into a deterministic nominal state driven by inputs and a stochastic error state affected by disturbances. Satisfaction of original chance constraints is guaranteed by tightening nominal constraints probabilistically with respect to additive disturbances and robustly with respect to measurement noise. The resulting data-driven receding horizon optimal control problem is lightweight, recursively feasible, and renders the closed loop input-to-state stable in the presence of both additive disturbances and measurement noise. We demonstrate the effectiveness of the proposed approach in a simulation example.
BibTeX
@article{kerz2021data,
  author = {Sebastian Kerz and Johannes Teutsch and Tim Brüdigam and Marion Leibold and Dirk Wollherr},
  title = {{Data-Driven Tube-Based Stochastic Predictive Control}},
  journal = {IEEE Open Journal of Control Systems},
  year = {2021},
  doi = {10.48550/arxiv.2112.04439},
}
Figure