Summary
In scenarios like online advertising markets and financial exchanges, autonomous, self-interested learning agents engage in strategic interactions via a shared platform. Platforms typically opt for a passive role, providing the shared infrastructure necessary for the operation of the multi-learner environment, while keeping learning procedures decentralized. Fully centralized systems can potentially yield superior outcomes by optimizing shared objectives like social welfare, but they are seldom chosen.
The goal of this project is to bridge the gap between these two extremes by establishing the theoretical foundations of semi-centralized platforms (SCP). SCPs aim to combine the best attributes of both centralized and decentralized systems, enabling next-generation platforms to operate efficiently at scale with the flexibility of decentralized learning, while also being able to steer learning agents towards desirable objectives. Using tools from online learning and computational game theory, we will develop innovative techniques to determine when and how platforms should actively influence the actions of learning agents. In this endeavor, we will i) develop a better understanding of the learning dynamics of traditional platforms; ii) explore methods to overcome well-known computational challenges that hinder the convergence of multi-learner systems towards shared objectives; and iii) extend fundamental game-theoretic models to realistic settings.
This research will pave the way for practical applications on real-world platforms and address pressing concerns related to fairness and accountability in their outcomes, which are expected to become even more significant as machine-learning algorithms gain wider adoption.
The goal of this project is to bridge the gap between these two extremes by establishing the theoretical foundations of semi-centralized platforms (SCP). SCPs aim to combine the best attributes of both centralized and decentralized systems, enabling next-generation platforms to operate efficiently at scale with the flexibility of decentralized learning, while also being able to steer learning agents towards desirable objectives. Using tools from online learning and computational game theory, we will develop innovative techniques to determine when and how platforms should actively influence the actions of learning agents. In this endeavor, we will i) develop a better understanding of the learning dynamics of traditional platforms; ii) explore methods to overcome well-known computational challenges that hinder the convergence of multi-learner systems towards shared objectives; and iii) extend fundamental game-theoretic models to realistic settings.
This research will pave the way for practical applications on real-world platforms and address pressing concerns related to fairness and accountability in their outcomes, which are expected to become even more significant as machine-learning algorithms gain wider adoption.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/101165466 |
| Start date: | 01-10-2024 |
| End date: | 30-09-2029 |
| Total budget - Public funding: | 1 316 544,00 Euro - 1 316 544,00 Euro |
Cordis data
Original description
In scenarios like online advertising markets and financial exchanges, autonomous, self-interested learning agents engage in strategic interactions via a shared platform. Platforms typically opt for a passive role, providing the shared infrastructure necessary for the operation of the multi-learner environment, while keeping learning procedures decentralized. Fully centralized systems can potentially yield superior outcomes by optimizing shared objectives like social welfare, but they are seldom chosen.The goal of this project is to bridge the gap between these two extremes by establishing the theoretical foundations of semi-centralized platforms (SCP). SCPs aim to combine the best attributes of both centralized and decentralized systems, enabling next-generation platforms to operate efficiently at scale with the flexibility of decentralized learning, while also being able to steer learning agents towards desirable objectives. Using tools from online learning and computational game theory, we will develop innovative techniques to determine when and how platforms should actively influence the actions of learning agents. In this endeavor, we will i) develop a better understanding of the learning dynamics of traditional platforms; ii) explore methods to overcome well-known computational challenges that hinder the convergence of multi-learner systems towards shared objectives; and iii) extend fundamental game-theoretic models to realistic settings.
This research will pave the way for practical applications on real-world platforms and address pressing concerns related to fairness and accountability in their outcomes, which are expected to become even more significant as machine-learning algorithms gain wider adoption.
Status
SIGNEDCall topic
ERC-2024-STGUpdate Date
23-11-2024
Images
No images available.
Geographical location(s)
