Descrizione del progetto
Niente «senno di poi» grazie ad una innovativa ottimizzazione del processo decisionale
Il processo decisionale in ingegneria, come nella vita reale, deve affrontare l’incertezza intrinseca che deriva da diverse fonti, tra cui l’accesso a dati osservabili limitati, il rumore della misurazione e il cosiddetto errore dell’operatore. Esistono numerose tecniche di ottimizzazione per tenere conto dell’incertezza con diverse ipotesi di partenza, tuttavia, man mano che i set di dati diventano più grandi e più complessi, aumenta anche l’incertezza. Il progetto TRUST, finanziato dall’UE, sta sviluppando un nuovo paradigma di controllo basato sul campo emergente dell’ottimizzazione distribuzionalmente solida, prendendo in considerazione tutte le distribuzioni di probabilità coerenti con le informazioni precedentemente fornite per affrontare le nuove sfide. I risultati includeranno un nuovo linguaggio di modellazione disponibile come software open-source.
Obiettivo
"Recent developments in sensing and communication technology offer unprecedented opportunities by ubiquitously collecting data at high detail and at large scale. Utilization of data at these scales, however, poses a major challenge for control systems, particularly in view of the additional inherent uncertainty that data-driven control signals introduce to systems behavior. In fact, this effect has not been well understood to this date, primarily due to the missing link between data analytics techniques in machine learning and the underlying physics of dynamical systems.
I address this issue by proposing a novel control design paradigm embracing ideas from the emerging field of distributionally robust optimization (DRO). DRO is a decision-making model whose solutions are optimized against all distributions consistent with given prior information. Recent breakthrough work, among others by the PI of this proposal, has shown that many DRO models can be solved in polynomial time even when the corresponding stochastic models are intractable. DRO models also offer a more realistic account of uncertainty and mitigate the infamous ""post-decision disappointment"" of stochastic models.
This proposal lays the theoretical foundation for distributionally robust control and aims to make progress along four directions. (i) Decision-dependent ambiguity: I introduce the concept of ""invariant ambiguity sets"" to encompass the dynamic evolution of uncertainty. (ii) Dynamic programming: I establish a dynamic programming characterization of the proposed DRO models and provide tractable approximation schemes along with rigorous theoretical bounds. (iii) Safe and memory-efficient learning: Leveraging modern tools from kernel methods and online-optimization, I propose tractable, yet provably reliable, synthesis tools. (iv) I plan to develop a tailor-made modeling language and open-source software to make distributionally robust control methods accessible in industry-size applications."
Campo scientifico
- natural sciencescomputer and information sciencessoftware
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringcontrol systems
- natural sciencesmathematicsapplied mathematicsdynamical systems
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
Programma(i)
Argomento(i)
Meccanismo di finanziamento
ERC-STG - Starting GrantIstituzione ospitante
2628 CN Delft
Paesi Bassi