When: 11 Avril 2025, 11h00 – 12h00

Where: Amphi 5, Palaiseau Campus (Télécom Paris)

Speaker: Assoc. Prof. Andrea Araldo (Télécom SudParis – SAMOVAR)

Title: « Optimally deceiving a learning leader in Stackelberg games »

Abstract:

I will present the setting and some interesting proofs contained in the NeurIPS 2020 paper « Optimally deceiving a learning leader in Stackelberg games ».

To ground the discussion in a concrete context, I will discuss the paper is a demand-side management scenario.

The regulator (e.g., the government – acting as a leader) proposes incentives to users (followers), to convince them to switch to sustainable alternatives (e.g., when choosing products or services). The regulator optimizes such incentives with the goal of maximizing social welfare. Users choose alternatives so as to maximize their utility function. The regulator has only partial information about the users’ true utility function. Therefore, users can can manipulate the regulator, by choosing alternatives according to a « fake » utility function, different than their true one. By doing so, users can mislead the regulator and receive more favourable incentives than they would deserve. The paper shows analytically under which conditions users can basically mislead the regulator as much as they want.