Efficient Automatic Tuning for Data-driven Model Predictive Control via Meta-Learning

TL;DR

This work tackles the inefficiency and instability of AutoMPC tuning caused by random initialization in Bayesian Optimization when facing new tasks. It introduces Portfolio, a meta-learning approach that warmstarts BO with a fixed, diverse set of configurations learned from prior tasks, yielding faster convergence and more robust tuning. Empirical results across 12 datasets (11 simulations and 1 underwater robot) show improved model identification and downstream control performance, especially for in-distribution tasks, with insights into how portfolio size affects stability. The study highlights practical benefits for resource-constrained robotics deployment and outlines future directions for adaptive portfolio selection and handling out-of-distribution data.

Abstract

AutoMPC is a Python package that automates and optimizes data-driven model predictive control. However, it can be computationally expensive and unstable when exploring large search spaces using pure Bayesian Optimization (BO). To address these issues, this paper proposes to employ a meta-learning approach called Portfolio that improves AutoMPC's efficiency and stability by warmstarting BO. Portfolio optimizes initial designs for BO using a diverse set of configurations from previous tasks and stabilizes the tuning process by fixing initial configurations instead of selecting them randomly. Experimental results demonstrate that Portfolio outperforms the pure BO in finding desirable solutions for AutoMPC within limited computational resources on 11 nonlinear control simulation benchmarks and 1 physical underwater soft robot dataset.

Figures (5)

• Figure 1: Overview of AutoMPC with Portfolio. Orange rectangular boxes refer to the input data and the blue rounded boxes refer to the output component in each process. (a) We build the Portfolio in a similar way that feurer-autosklearn does for AutoML. We first perform AutoMPC model tuning without meta-learning for each meta dataset $D_i$, obtaining the corresponding optimal model configuration $h_i$. These configurations form the candidate set $C$. Then, we evaluate the performance of $C$ on $D_{meta}$ and construct a performance matrix where $S_{D_i,h_j}$ represents the score of configuration $h_j$ on meta dataset $D_i$. Finally, we employ the Greedy Portfolio algorithm proposed by feurer-autosklearn to obtain the Portfolio $P$. (b) We utilize $P$ as the initial configurations for BO in AutoMPC model tuning. It will return the best model configuration $h^*$ and correspondent surrogate dynamics model $f_{h^*}$, which will be used for control tuning in AutoMPC and helps to achieve the optimal synthesized MPC.
• Figure 2: Comparison between Portfolio with size 10 and Pure BO. We evaluate the model tuning performance on Walker2dSmall and HopperGravityOneAndHalf datasets within 100 iterations. The blue lines represent the tuning curve for Pure BO, while the red lines denote the tuning curve for Portfolio with size 10. Portfolio can lead to a faster convergence of model tuning and outperform pure BO within limited time.
• Figure 3: Comparison among different Portfolio sizes. We evaluate the model tuning performance on HopperSmallTorso dataset with pure BO (Portfolio 0) and Portfolio with sizes $5, 10, 15, 20$ on 5 independent runs, plotted in grey. The median and best result are highlighted in blue and green, respectively. In this case, Portfolio consistently outperforms the pure BO regardless of the size. The tuning process becomes more stable as the Portfolio size increases.
• Figure 4: Gains among different Portfolio sizes. We compare the gains of Portfolio with sizes $5, 10, 15, 20$ over the pure BO on 6 datasets. The best Portfolio size varies for different datasets and an inappropriate Portfolio size may worsen the performance.
• Figure 5: Tuning curves for Portfolio with size 10 and Pure BO.