Thèse Learning Algorithms For Designing Undeceivable Policies To Foster Sustainable Behavior H/F Doctorat.Gouv.Fr
- Paris - 75
- CDD
- Service public d'état
Les missions du poste
Task3: Sequential decision-making formulation. We will model the agent as a sequential decision-maker willing to maximize their cumulative true utility. Via a Markov Decision Process formulation, we will compute the theoretically optimal policy of the agent, i.e., the sequence of choices to select in order to maximize the cumulative true utility of the agent. We proved in our preliminary work[Da25] that this policy is superior to adopting a fake utility. In
the online advertisement domain, agents are modeled as sequential decision-makers. While observing millions of transactions, the regulator sets prices via bandit-like algorithms [Mo15, Ro13] or based on perfect knowledge of the agent utility distribution. [Dr19] However, such methods are not usable for PDSMS, in which observed choices are much fewer. Therefore, we will apply more parsimonious techniques, rooted in Economics, [Se24,Ch21] to compute signals, such as personalized prices and incentives, coupled in an original way with the sequential decision-maker nature of the agents.
Task4: Lab experiments with~30 participants, recruited among students and colleagues, will be organized. Indeed, by resorting to Experimental Economics[Uk17] we can test our theoretical findings and conjectures in a human setting. Participants will take the role of agents and interact with a simulated regulator in a fictitious scenario with congestion, e.g., daily commute scenario. Real small monetary incentives will be provided to the participants, to link their choices to real consequences, according to the Induced Value Theory.[Sm62] We will assess the impact, on the Cost of Deception (CoD) and on the intensity of deceptive behavior, of the regulator's learning speed (testing more or less efficient preference elicitation techniques), data borrowing, congestion severity.
Task5: Mechanism Design. We will describe PDSMS within the framework of Mechanism Design.[Bo15] In the Vickrey-Clarke-Groves (VCG) Vi61,Cl71,Gr73 mechanism terminology, regulator & agents correspond to players, agents' choices to the outcome, signals to the price function. We will assess fundamental trade-offs between social welfare maximization and strategy-proofness (i.e., robustness to deception). Our original contribution is that we will exploit
findings from Task2, to artificially limit the goodness of fit of the preference elicitation techniques run by the regulator, so as to limit the deception power of the agents. To this aim, we will consider preferring simple statistical models with few degrees of freedom to complex ones, introducing noise, adopting small learning rates. With the same aim, we will study if the regulator can reduce vulnerability to deceptive behavior by limiting the amount of personalization, in order to be less sensitive to agents' manipulations. In the extreme, a non-personalized policy, such as a random coefficient[Da19], is extremely robust to deceptive agents. However, it cannot adapt to the peculiarity of the agents. We will thus avoid such extreme cases. We will thus study in this task the fundamental trade-off between robustness to deception and social welfare maximization, adjusting the right amount of personalization
Le profil recherché
Duration: 3 years
Requirements: Excellent mathematical modeling and analytical skills, good programming skills
To apply: Please send your CV, an explanation of 5 lines explaining why you are the best fit for this position (with factual non-vague or generic elements), all the marks of your BSc and MSc level courses; Sending your ranking is not mandatory (but it is a big plus). Send all this material to ****@****.**
Bienvenue chez Doctorat.Gouv.Fr
Our objective is to build a methodological framework to understand under which conditions Personalized Demand-Side Mitigation Strategies (PDSMS) are robust to deceptive behavior and how to improve such robustness. Expected results are: (i) Analysis of the extent to which agents can benefit from deceptive behavior, (ii) Assessment the cost of deception suffered by agents to mislead the regulator (iii) Design of mechanisms to deter deceptive behavior and (iv) Precise characterization of the trade-off between social welfare and vulnerability to deception, to understand up to which point the regulator might accept to be manipulated by the agents, while maximizing social welfare. SCIENTIFIC CONTENT (research activities presented in approximate chronological order)
Task1: Game Theory formulation. We will formulate a repeated leader-follower game, [Br88,Ke23] with the leader (regulator) and, initially, one follower (agent) only, who can be deceptive, adopting a fake utility to get a more favorable signal. We will then consider equilibrium between multiple followers, subject to congestion effects, as in [Ro19, Ta22]. Compared to [Ro19, Ta22], we will additionally include deceptive agents. We conjecture that congestion can make deception more difficult, as its effectiveness is limited by the presence of the other agents. We will study how social welfare from Personalized Demand-Side Mitigation Strategies (PDSMS) is impacted by different fractions of deceptive agents.
Task2: Preference elicitation and Cost of Deception. In Task1 we assume that the regulator knows the utility function followed by agents (either the true or the fake one). Here, we will instead explicitly model the statistical methods with which the regulator can learn the agent's utility function. We will adopt state-of-the-art preference elicitation techniques, generally used in marketing.[Be19,Da19,Go20,Zi04,Zi03,La04] Such methods may be slow and collect limited information. Such limits are usually considered an inconvenience for the regulator. Instead, we conjecture that in the particular case of PDSMS, such limits are a protection against deceptive users. Indeed, an agent needs to suffer a high Cost of Deception (CoD), for many iterations, to be able to manipulate a regulator that learns their characteristics slowly. We will systematically study an interesting counterintuitive phenomenon, which has been observed in other scenarios,[Xu21] and that we conjecture would more strongly hold in PDSMS: deliberately reducing the complexity or the precision of the preference elicitation techniques run by the regulator may be desirable, as it reduces the risk of overfitting to agents' manipulation, and increases the CoD. We will also study the impact on the CoD of data borrowing,[Ba00] a technique that improves the elicitation of an agent's preferences by exploiting not only their observed behavior, but also the behavior of similar agents. Our conjecture is that data borrowing can also increase CoD and thus the robustness of PDSMS.
Publiée le 20/10/2026 - Réf : a96e1088a9eb4c7bb5ae808155d2b6b0