Modern production systems require decision-support tools capable of determining optimal operating conditions while accounting for multiple, conflicting factors and constraints. This is particularly critical when balancing sustainability objectives with production targets, requiring a thorough analysis of possible optimal scenarios. Cooperation between human cognitive capabilities and advanced learning and control modules is central to the Industry 5.0 paradigm, leading to the development of frameworks such as Human–Cyber–Physical Systems (HCPS), which aim to enhance human capabilities without replacing them. The decision-maker plays a crucial role in identifying the best compromise and providing real-time feedback, based on information derived from data analysis, while adapting control modules according to their instructions. In this context, Multi-Objective Optimization (MOO) offers a rigorous framework to explore trade-offs and support reliable decisions. This thesis develops an enhanced a posteriori approach by adopting a Pareto front (PF) surrogate with robust and balanced approximation error guarantees, enabling informed choices and an efficient representation of the complete set of optimal alternatives. The methodology introduces a linear and invertible transformation to convert the problem of identifying the PF into that of identifying a black-box function, in which the solution set computed by scalarization corresponds to the available dataset. First, an algorithm for the bi-objective case, named ROBBO (RObust and Balanced Bi-objective Optimization), is introduced. It computes a continuous approximation of the entire PF with guaranteed accuracy. Moreover, it provides necessary and sufficient theoretical guarantees on the dataset to fulfill user-defined requirements. This strategy is rigorously shown to be superior to classical methods in terms of sampling complexity, offering prior information on the worst-case number of Pareto-optimal solutions required. The approach is showcased in a constrained path-following problem for a two-axis positioning system and in a steady-state optimization problem for a Continuous-Flow Stirred Tank Reactor. Extensions are presented to handle disconnected PFs by combining a modified Pascoletti–Serafini scalarization with a systematic detection of gaps. The approach is then generalized to the multi-objective case, RObust and Balanced Multi-Objective Optimization (ROBMO). A new transformation is introduced to convert the MOO problem into a black-box function estimation with unknown domain. New guarantees are established regarding the PF approximation error and the upper bound on the number of samples required in the worst case. The algorithm is further adapted to estimate the boundaries of the PF (i.e., the domain of the function), and its effectiveness is demonstrated through benchmarks in problems with more than two objectives. The methodology is validated within an HCPS, deployed on an industrial large-scale manufacturing plant for Medium-Density Fiberboard production. A Multiple-Criteria Decision Support System, interconnected with a digital twin and a model predictive control, and based on the ROBBO algorithm, is designed to assist operators in balancing production and sustainability by suggesting optimal process parameters. The system supports what-if analyses and transmits realized operating points to the control layer. The scheme has been implemented and tested in production, representing one of the few documented examples of such systems deployed in large-scale industrial plants. Finally, the ROBBO algorithm is extended to deal with objective functions featuring parametric uncertainty, where the PF constitutes a bundle of functions, and scenario optimization is used to derive probabilistic accuracy guarantees for the approximation using only a limited number of measures of the uncertain element. The research was conducted in the context of the European project E2COMATION, “Life cycle optimization of industrial energy efficiency by a distributed control and decision-making automation platform”, which aims to improve industrial energy efficiency by integrating data analysis with advanced optimization techniques, addressing both production and environmental perspectives.
I sistemi di produzione moderni richiedono strumenti di supporto alle decisioni in grado di determinare le condizioni operative ottimali, tenendo conto di molteplici vincoli e fattori tra loro contrastanti. Ciò è particolarmente critico quando si confrontano gli obiettivi di sostenibilità con i target di produzione, richiedendo un'analisi approfondita degli scenari possibili. La collaborazione tra le capacità cognitive umane e i moduli avanzati di apprendimento e controllo è centrale per il paradigma Industria 5.0, portando allo sviluppo di framework come gli Human–Cyber–Physical Systems (HCPS), che mirano a potenziare le capacità umane senza sostituirle. Il decision-maker gioca un ruolo cruciale nell'identificare il miglior compromesso e nel fornire feedback in tempo reale, basandosi sulle informazioni derivanti dalle misurazioni e adattando al contempo i moduli di controllo. In questo contesto, l’ottimizzazione multi-obiettivo (MOO) offre un metodo rigoroso per esplorare i trade-off e supportare in modo affidabile le decisioni. Questa tesi sviluppa un approccio a posteriori aumentato, sfruttando un modello del fronte di Pareto (PF) con garanzie di errore di approssimazione robuste e bilanciate, consentendo scelte informate e una rappresentazione efficiente dell'insieme completo delle alternative ottimali. La metodologia introduce una trasformazione lineare e invertibile per convertire il problema dell'identificazione del PF in quello dell'identificazione di una funzione black-box, in cui l'insieme delle soluzioni, calcolate tramite metodi di scalarizzazione, corrisponde al dataset disponibile. In primo luogo, viene introdotto un algoritmo per il caso bi-obiettivo, denominato ROBBO (RObust and Balanced Bi-objective Optimization). Esso calcola un'approssimazione continua dell'intero PF con accuratezza garantita. Inoltre, fornisce garanzie teoriche necessarie e sufficienti sul dataset per soddisfare i requisiti definiti dall'utente. Viene dimostrato rigorosamente che questa strategia è superiore ai metodi classici in termini di numero di campioni, offrendo informazioni a priori sul numero di soluzioni Pareto ottime richieste nel caso peggiore. L'approccio viene illustrato in un problema di constrained path-following per un sistema di posizionamento a due assi e in un problema di steady-state optimization per un Continuous-Flow Stirred Tank Reactor. Vengono presentate estensioni per gestire PF disconnessi combinando la scalarizzazione di Pascoletti–Serafini con un rilevamento sistematico dei gap del PF. L'approccio viene poi generalizzato al caso multi-obiettivo, RObust and Balanced Multi-Objective Optimization (ROBMO). Viene introdotta una nuova trasformazione per convertire il problema MOO in una stima di funzione black-box con dominio non noto a priori. Vengono stabilite nuove garanzie riguardanti l'errore di approssimazione del PF e il limite superiore del numero di campioni necessari nel caso peggiore. L'algoritmo è ulteriormente adattato per stimare i bordi del PF (ovvero, il dominio della funzione), e la sua efficacia è dimostrata attraverso benchmark in problemi con più di due obiettivi. La metodologia è validata all'interno di un HCPS, implementato in un impianto manifatturiero industriale di larga scala per la produzione di Medium-Density Fiberboard. Un Multiple-Criteria Decision Support System, interconnesso con un digital twin e un model predictive control, e basato sull'algoritmo ROBBO, è progettato per assistere gli operatori nel bilanciare produzione e sostenibilità suggerendo parametri di processo ottimali. Il sistema supporta analisi what-if e trasmette i punti di funzionamento selezionati al controllore. Lo schema è stato implementato e testato in produzione, rappresentando uno dei pochi esempi documentati di tali sistemi implementati in impianti industriali di larga scala. Infine, l'algoritmo ROBBO viene esteso per gestire funzioni obiettivo caratterizzate da incertezza parametrica, dove il PF costituisce un fascio di funzioni, e l'ottimizzazione a scenari viene utilizzata per derivare garanzie di accuratezza probabilistica sull'errore di approssimazione, impiegando un numero limitato di misurazioni dei parametri incerti. La ricerca è stata condotta nel contesto del progetto europeo E2COMATION, "Life cycle optimization of industrial energy efficiency by a distributed control and decision-making automation platform ", che mira a migliorare l'efficienza energetica in ambito industriale, integrando l'analisi dei dati con tecniche avanzate di ottimizzazione, affrontando le prospettive sia produttive che ambientali.
Robust and balanced multi-objective optimization: methodology and applications
BOFFADOSSI, ROBERTO
2026
Abstract
Modern production systems require decision-support tools capable of determining optimal operating conditions while accounting for multiple, conflicting factors and constraints. This is particularly critical when balancing sustainability objectives with production targets, requiring a thorough analysis of possible optimal scenarios. Cooperation between human cognitive capabilities and advanced learning and control modules is central to the Industry 5.0 paradigm, leading to the development of frameworks such as Human–Cyber–Physical Systems (HCPS), which aim to enhance human capabilities without replacing them. The decision-maker plays a crucial role in identifying the best compromise and providing real-time feedback, based on information derived from data analysis, while adapting control modules according to their instructions. In this context, Multi-Objective Optimization (MOO) offers a rigorous framework to explore trade-offs and support reliable decisions. This thesis develops an enhanced a posteriori approach by adopting a Pareto front (PF) surrogate with robust and balanced approximation error guarantees, enabling informed choices and an efficient representation of the complete set of optimal alternatives. The methodology introduces a linear and invertible transformation to convert the problem of identifying the PF into that of identifying a black-box function, in which the solution set computed by scalarization corresponds to the available dataset. First, an algorithm for the bi-objective case, named ROBBO (RObust and Balanced Bi-objective Optimization), is introduced. It computes a continuous approximation of the entire PF with guaranteed accuracy. Moreover, it provides necessary and sufficient theoretical guarantees on the dataset to fulfill user-defined requirements. This strategy is rigorously shown to be superior to classical methods in terms of sampling complexity, offering prior information on the worst-case number of Pareto-optimal solutions required. The approach is showcased in a constrained path-following problem for a two-axis positioning system and in a steady-state optimization problem for a Continuous-Flow Stirred Tank Reactor. Extensions are presented to handle disconnected PFs by combining a modified Pascoletti–Serafini scalarization with a systematic detection of gaps. The approach is then generalized to the multi-objective case, RObust and Balanced Multi-Objective Optimization (ROBMO). A new transformation is introduced to convert the MOO problem into a black-box function estimation with unknown domain. New guarantees are established regarding the PF approximation error and the upper bound on the number of samples required in the worst case. The algorithm is further adapted to estimate the boundaries of the PF (i.e., the domain of the function), and its effectiveness is demonstrated through benchmarks in problems with more than two objectives. The methodology is validated within an HCPS, deployed on an industrial large-scale manufacturing plant for Medium-Density Fiberboard production. A Multiple-Criteria Decision Support System, interconnected with a digital twin and a model predictive control, and based on the ROBBO algorithm, is designed to assist operators in balancing production and sustainability by suggesting optimal process parameters. The system supports what-if analyses and transmits realized operating points to the control layer. The scheme has been implemented and tested in production, representing one of the few documented examples of such systems deployed in large-scale industrial plants. Finally, the ROBBO algorithm is extended to deal with objective functions featuring parametric uncertainty, where the PF constitutes a bundle of functions, and scenario optimization is used to derive probabilistic accuracy guarantees for the approximation using only a limited number of measures of the uncertain element. The research was conducted in the context of the European project E2COMATION, “Life cycle optimization of industrial energy efficiency by a distributed control and decision-making automation platform”, which aims to improve industrial energy efficiency by integrating data analysis with advanced optimization techniques, addressing both production and environmental perspectives.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Boffadossi_DEF_perSTAMPA.pdf
embargo fino al 17/05/2027
Licenza:
Tutti i diritti riservati
Dimensione
45.56 MB
Formato
Adobe PDF
|
45.56 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/382837
URN:NBN:IT:POLIMI-382837