Grant: Discovery Grants Program - Individual

Introduction My work focuses on online learning and optimization, where an agent interacts sequentially with a static, stochastic, or possibly adversarial environment. During each interaction, the agent gathers some feedback, with the end goal of optimizing some unknown objective under minimal a priori assumptions. Depending on the specific nature of each problem, finite-time convergence rates for cumulative regret or optimization error (also known as simple regret) can be derived, and it is possible to design algorithms with provably optimal performance guarantees. This formalism captures many fundamental settings of great practical interest, including healthcare, economics, and internet technologies. Long-Term Goals - Determining the theoretical hardness of the most interesting open online learning and optimization problems - Developing algorithms whose performances provably match the best achievable limit - Expand the classic viewpoint of worst-case guarantees to the more challenging instance-dependent approach, as motivated by the needs of most modern applications - Maintain a consistent output in the training of highly qualified persons in the theory of online learning and optimization Short-Term Goals: Research - Bilateral Trade: Determining how efficiently a platform intermediating between a seller and a buyer can learn the price that maximizes social welfare in an adversarial setting with discretized prices, with applications to ride-sharing apps - Supply Chains: Investigating if it is possible to converge to an optimal strategy (or some Stackelberg equilibrium) when suppliers interact sequentially with retailers that, in turn, make transactions with the market. It would be of great practical interest to understand how these dynamics behave under different assumptions about the players' behavior (e.g., acting selfishly or cooperating in a vertical-integration regime) - Online Optimization: Fleshing out the optimal theoretical guarantees of lower bounding algorithms for high-dimensional non-convex objectives (e.g., Piyavskii-Shubert) and low-dimensional convex ones (e.g., Halving), focusing on instance-dependent bounds in zeroth order optimization, as all these variants of optimization problems have their place in the landscape of applications - Space Operations: In a recent collaboration with a partner of the European Space Agency (EUMETSAT), we obtained remarkable results, improving the existing offset-optimization technology by more than 60%. Concretely, this translates into saving hundreds of thousands of dollars of data. I am committed to continuing my exploration of the novel field of machine learning for space operations and bringing my experience to the Canadian Space Agency More on HQP Training I passionately participated in cosupervising and tutoring students during my Ph.D. and Post-Doc years. I look forward to mentoring new students in my first professor position and promote a fair, diverse, and inclusive environment.