In this technical communique, we present a sim2real approach for data-driven controller tuning, utilizing a digital twin to generate input–output data and suitable controllers around nominal parameter values. We establish a direct inverse supervised learning framework using advanced neural network architectures, including the WaveNet sequence model, to learn a tuning rule that maps input–output data to controller parameters. This approach automates controller re-calibration by meta-learning the tuning rule through inverse supervised learning, effectively avoiding human intervention via a machine learning model. The advantages of this methodology are demonstrated through numerical simulations across various neural network architectures.

For many industrial processes, a digital twin is available, which is essentially a highly complex model whose parameters may not be properly tuned for the specific process. By relying on the availability of such a digital twin, this paper introduces a novel approach to data-driven control, where the digital twin is used to generate samples and suitable controllers for various perturbed versions of its parameters. A supervised learning algorithm is then employed to estimate a direct mapping from the data to the best controller to use. This map consists of a model reduction step, followed by a neural network architecture whose output provides the parameters of the controller. The data-to-controller map is pre-computed based on artificially generated data, but its execution once deployed is computationally very efficient, thus providing a simple and inexpensive way to tune and re-calibrate controllers directly from data. The benefits of this novel approach are illustrated via numerical simulations.

Parameter estimation is a pivotal task across various domains such as system identification, statistics, and machine learning. The literature presents numerous estimation procedures, many of which are backed by well-studied asymptotic properties. In the contemporary landscape, highly advanced digital twins (DTs) offer the capability to faithfully replicate real systems through proper tuning. Leveraging these DTs, data-driven estimators can alleviate challenges inherent in traditional methods, notably their computational cost and sensitivity to initializations. Furthermore, traditional estimators often rely on sensitive data, necessitating protective measures.

In this thesis, we consider data-driven and privacy-preserving approaches to parameter estimation that overcome many of these challenges.

The first part of the thesis delves into an exploration of modern data-driven estimation techniques, focusing on the two-stage (TS) approach. Operating under the paradigm of inverse supervised learning, the TS approach simulates numerous samples across parameter variations and employs supervised learning methods to predict parameter values. Divided into two stages, the approach involves compressing data into a smaller set of samples and the second stage utilizes these samples to predict parameter values. The simplicity of the TS estimator underscores its interpretability, necessitating theoretical justification, which forms the core motivation for this thesis. We establish statistical frameworks for the TS estimator, yielding its Bayes and minimax versions, alongside developing an improved minimax TS variant that excels in computational efficiency and robustness to distributional shifts. Finally, we conduct an asymptotic analysis of the TS estimator.

The second part of the thesis introduces an application of data-driven estimation methods, that includes the TS and neural network based approaches, in the design of tuning rules for PI controllers. Leveraging synthetic datasets generated from DTs, we train machine learning algorithms to meta-learn tuning rules, streamlining the calibration process without manual intervention.

In the final part of the thesis, we tackle scenarios where estimation procedures must handle sensitive data. Here, we introduce differential privacy constraints into the Bayes point estimation problem to protect sensitive information. Proposing a unified approach, we integrate the estimation problem and differential privacy constraints into a single convex optimization objective, thereby optimizing the accuracy-privacy trade-off. In cases where both observations and parameter spaces are finite, this approach reduces to a tractable linear program which is solvable using off-the-shelf solvers.

In essence, this thesis endeavors to address computational and privacy concerns within the realm of parameter estimation.

Skattning av parametrar utgör en fundamental uppgift inom en mängd fält, såsom systemidentifiering, statistik och maskininlärning. I litteraturen finns otaliga skattningsmetoder, utav vilka många understödjs av välstuderade asymptotiska egenskaper. Inom dagens forskning erbjuder noggrant kalibrerade digital twins (DTs) möjligheten att naturtroget återskapa verkliga system. Genom att utnyttja dessa DTs kan data-drivna skattningsmetoder minska problem som vanligtvis drabbar traditionella skattningsmetoder, i synnerhet problem med beräkningsbörda och känslighet för initialiseringvillkor. Traditionella skattningsmetoder kräver dessutom ofta känslig data, vilket leder till ett behov av skyddsåtgärder.

I den här uppsatsen, undersöker vi data-drivna och integritetsbevarande parameterskattningmetoder som övervinner många av de nämnda problemen. 

Första delen av uppsatsen är en undersökning av moderna data-drivna skattningtekniker, med fokus på två-stegs-metoden (TS). Som metod inom omvänd övervakad maskininlärning, simulerar TS en stor mängd data med ett stort urval av parametrar och tillämpar sedan metoder från övervakad inlärning för att förutsäga parametervärden. De två stegen innefattar datakomprimering till en mindre mängd, varefter den mindre mängden data används för parameterskattning. Tack vare sin enkelhet och tydbarhet lämpar sig två-stegs-metoden väl för teoretisk analys, vilket är uppsatsens motivering.

Vi utvecklar ett statistiskt ramverk för två-stegsmetoden, vilket ger Bayes och minimax-varianterna, samtidigt som vi vidareutvecklar minimax-TS genom en variant med hög beräkningseffektivitet och robusthet gentemot skiftade fördelningar. Slutligen analyserar vi två-stegs-metodens asymptotiska egenskaper. 

Andra delen av uppsatsen introducerar en tillämpning av data-drivna skattningsmetoder, vilket innefattar TS och neurala nätverk, i designen och kalibreringen av PI-regulatorer. Med hjälp av syntetisk data från DTs tränar vi maskininlärningsalgoritmer att meta-lära sig regler för kalibrering, vilket effektiverar kalibreringsprocessen utan manuellt ingripande. 

I sista delen av uppsatsen behandlar vi scenarion då skattningsprocessen innefattar känslig data. Vi introducerar differential-privacy-begränsningar i Bayes-punktskattningsproblemet för att skydda känslig information. Vi kombinerar skattningsproblemet och differential-privacy-begränsningarna i en gemensam konvex målfunktion, och optimerar således avvägningen mellan noggrannhet och integritet. Ifall både observations- och parameterrummen är ändliga, så reduceras problemet till ett lätthanterligt linjärt optimeringsproblem, vilket löses utan vidare med välkända metoder. 

Sammanfattningsvis behandlar uppsatsen beräkningsmässiga och integritets-angelägenheter inom ramen för parameterskattning.

Parameter estimation in statistics and system identification relies on data that may contain sensitive information. To protect this sensitive information, the notion of differential privacy (DP) has been proposed, which enforces confidentiality by introducing randomization in the estimates. Standard algorithms for differentially private estimation are based on adding an appropriate amount of noise to the output of a traditional point estimation method. This leads to an accuracy-privacy trade off, as adding more noise reduces the accuracy while increasing privacy. In this paper, we propose a new Unified Bayes Private Point (UBaPP) approach to Bayes point estimation of the unknown parameters of a data generating mechanism under a DP constraint, that achieves a better accuracy-privacy trade off than traditional approaches. We verify the performance of our approach on a simple numerical example.

Parameter estimation is an important sub-field in statistics and system identification. Various methods for parameter estimation have been proposed in the literature, among which the Two-Stage (TS) approach is particularly promising, due to its ease of implementation and reliable estimates. Among the different statistical frameworks used to derive TS estimators, the min-max framework is attractive due to its mild dependence on prior knowledge about the parameters to be estimated. However, the existing implementation of the minimax TS approach has currently limited applicability, due to its heavy computational load. In this paper, we overcome this difficulty by using a gradient boosting machine (GBM) in the second stage of TS approach. We call the resulting algorithm the Two-Stage Gradient Boosting Machine (TSGBM) estimator. Finally, we test our proposed TSGBM estimator on several numerical examples including models of dynamical systems.

One of the most important problems in system identification and statistics is how to estimate the unknown parameters of a given model. Optimization methods and specialized procedures, such as Empirical Minimization (EM) can be used in case the likelihood function can be computed. For situations where one can only simulate from a parametric model, but the likelihood is difficult or impossible to evaluate, a technique known as the Two-Stage (TS) Approach can be applied to obtain reliable parametric estimates. Unfortunately, there is currently a lack of theoretical justification for TS. In this paper, we propose a statistical decision-theoretical derivation of TS, which leads to Bayesian and Minimax estimators. We also show how to apply TS on models for independent and identically distributed samples, by computing quantiles of the data as a first step, and using a linear function as the second stage. The proposed method is illustrated via numerical simulations.