Explicit MPC: Hard constraint satisfaction under low precision arithmetic

Andrea Suardi, Stefano Longo, Eric C. Kerrigan, George A. ConstantinidesControl 2015 · DOI

TL;DR

This paper presents a novel method to ensure hard constraint satisfaction in explicit Model Predictive Control (MPC) when implemented on embedded systems using low-precision arithmetic. By analyzing and bounding the numerical errors introduced by finite-precision computation, they design a robust controller that guarantees performance even with limited computational resources.

What Problem Does This Paper Address?

Modern embedded systems are increasingly used for control applications, driven by the need for lower costs, faster execution, and reduced power consumption. This often necessitates the use of low-precision arithmetic (e.g., fixed-point) for computations. However, standard control design often assumes infinite numerical precision, leading to a disconnect between theoretical guarantees and practical implementation. Specifically, low-precision arithmetic can introduce numerical errors that compromise the constraint satisfaction guarantees of MPC, a powerful but computationally intensive control technique. This paper addresses the challenge of maintaining hard constraint satisfaction in explicit MPC when implemented on resource-constrained embedded platforms that employ finite-precision arithmetic.

Methods

The core of the proposed method is an iterative approach to design a robust explicit MPC controller. The process involves:

  1. Forward Error Analysis: The authors use tools from forward error analysis to quantify the maximum possible error introduced by the finite-precision implementation of the explicit MPC control law. This error is treated as an additive disturbance to the plant dynamics.
  2. Robust Controller Design: A robust controller is then designed to be resilient to this quantified disturbance. This involves solving a minimax optimization problem to find bounds on the state disturbance.
  3. Iterative Refinement: The process is iterative. The robust controller designed in one iteration is used to generate new, potentially tighter, error bounds. These new bounds are then used to design the next-iteration robust controller. This iterative refinement continues until a satisfactory trade-off between robustness and conservativeness is achieved.
  4. Point Location and Function Evaluation: For explicit MPC, the control action is determined by first locating the current state within a pre-computed polyhedral partition of the state space and then evaluating a piecewise affine control law. The paper details how to handle errors in both the point location (using a binary search tree) and function evaluation steps due to finite precision.
  5. MILP/LP for Error Bound Computation: To compute the maximum and minimum bounds of the state disturbance, the authors formulate mixed-integer linear programming (MILP) problems. To reduce computational complexity, they propose solving a relaxed linear programming (LP) problem, which provides a slightly more conservative but acceptable bound.

Datasets

Not discussed in the paper. The paper focuses on a theoretical framework and experimental validation using a simulated inverted pendulum system.

Key Insights and Ideas

  • Quantifying Finite-Precision Errors is Crucial: The paper highlights that simply ignoring numerical errors from low-precision arithmetic is not an option for guaranteeing constraint satisfaction. A systematic way to bound these errors is essential.
  • Iterative Robustness Design: The iterative approach to designing robust controllers, where each iteration tightens the error bounds and designs a controller robust to those bounds, is a key contribution. This allows for a more precise and less conservative design compared to a single-shot approach.
  • Error as a Bounded Disturbance: Treating the finite-precision error as a bounded additive disturbance to the plant dynamics is a powerful abstraction that enables the application of robust control techniques.
  • Trade-off between Conservativeness and Precision: The iterative process implicitly manages the trade-off between ensuring constraint satisfaction (requiring more conservative bounds) and achieving better performance (allowing less conservative bounds).
  • FPGA Implementation Validation: The practical implementation on an FPGA board demonstrates the feasibility and effectiveness of the proposed method in a real-world embedded control scenario.

Key Results

  • The proposed method successfully guarantees hard constraint satisfaction for an explicit MPC controller implemented with low-precision arithmetic (as low as 8-bit fixed-point).
  • Experimental results on an inverted pendulum system showed that a nominal controller implemented with fixed-point arithmetic led to constraint violations, while the robustly designed controller maintained constraint satisfaction.
  • The iterative design process led to a reduction in the number of regions and hyperplanes in the explicit MPC controller, resulting in lower implementation complexity, reduced memory usage, and potentially lower energy consumption.
  • Test ID 4, a controller designed using the proposed method, guaranteed constraint satisfaction with minimal conservativeness, whereas Test ID 5, designed with less conservative bounds, failed to guarantee satisfaction.

Limitations and Future Work

  • Computational Complexity of Error Bound Calculation: While the paper proposes an LP relaxation to mitigate the complexity of MILP for error bound computation, it acknowledges that MILP can be computationally demanding, especially with a large number of bits for fraction length.
  • Convergence of Error Bounds: The paper notes that the numerical error bounds do not necessarily converge in the iterative process, and the behavior can be unpredictable. However, the iterative approach still allows for finding a suitable controller.
  • Future Work: The authors suggest extending this technique to other controller design methods. They also mention that the proposed controller design and HIL test procedure have been applied to various plants in different configuration scenarios, implying potential for broader applicability.

Summary updated

BibTeX
@inproceedings{suardi2015explicit,
  author = {Andrea Suardi and Stefano Longo and Eric C. Kerrigan and George A. Constantinides},
  title = {{Explicit MPC: Hard constraint satisfaction under low precision arithmetic}},
  booktitle = {Control},
  year = {2015},
  doi = {10.1016/j.conengprac.2015.12.005},
}

Loading related content...

Figure