Alexandre Rio

Algorithms are increasingly driving decision-making in the real world, impacting people's lives and society at large. This has unsurprisingly sparked ethical concerns, such as privacy and fairness, which are becoming ever more urgent to address. Tackling ethical issues in this context is challenging and often conflicts with the primary performance-oriented goals. However, increasingly stringent legal regulations and strategic interests provide strong incentives for technology organizations to invest in the development of trustworthy solutions.

In this context, this thesis contributes to the development of algorithms that integrate ethical issues, especially privacy and fairness, without sacrificing performance in sequential decision-making --- a setting that covers a wide range of problems where a series of decisions is taken over time. This effort is addressed through the lens of three sequential decision-making problems with many practical applications.

The thesis first addresses the problem of best arm identification in multi-agent multi-armed bandits, a framework used to optimize distributed large-scale systems. In this setting, we identify inter-agent communications as a critical privacy vulnerability. We therefore develop methods that rely on private communications without compromising the benefits of multi-agent collaboration.

The second part of this thesis deals with privacy-preserving deep reinforcement learning. While reinforcement learning is increasingly deployed in real-world scenarios with sensitive data, the field lacks scalable private solutions. This thesis aims to fill this gap by proposing deep reinforcement learning algorithms with formal differentially privacy guarantees. It first proposes a private model-based approach for the offline setting, before investigating theoretically-grounded differentially private policy gradient methods. 

Finally, the thesis investigates the integration of producer fairness in bundle recommendation, where users are served with sets of compatible and complementary items, and fairness of exposure is measured over entire recommendation sequences. After formalizing this challenging task, both exact and heuristic solutions are introduced, effectively balancing fairness of exposure and bundle quality.