Speaker: Alessio Russo
17 Dicembre 2025 | 13:00
DEIB, Sala Conferenze "E. Gatti" (Ed. 20)
Contatti: Prof. Marcello Restelli
Sommario
On December 17, 2025, at 1:00 pm the seminar on "A Tale of Active Testing: From Best Arm Identification to In-Context Pure Exploration" will take place at DEIB Conference Room "Emilio Gatti" (Building 20).Many sequential decision problems boil down to pure exploration: how many experiments do we need to confidently identify the underlying true hypothesis? I will start from the classical Best Arm Identification (BAI) problem in multi-armed bandits, using it to introduce the key ideas behind instance-dependent sample complexity lower bounds and how they shape optimal exploration strategies. I will then move to Best Policy Identification (BPI) in Markov Decision Processes, where the problem structure makes the design of efficient pure exploration algorithms more subtle.
In the second part of the talk, I will connect these ideas to show how we can use in-context learning (ICL) for learning optimal exploration strategies. Transformers meta-trained across families of pure exploration tasks can learn to implement sophisticated exploration strategies in-context: given a new task, they adaptively allocate measurements and stop when confident enough to make a recommendation of what is the true underlying hypothesis, all without parameter updates. I will present In-Context Pure Explorer (ICPE), a Transformer-based architecture that performs Active Sequential Hypothesis Testing (ASHT) in-context: at inference time, ICPE gathers evidence, identifies the ground-truth hypothesis, and competes with instance-dependent algorithms on BAI and generalized search tasks, while requiring only a forward pass and avoiding explicit optimization at test time.
Biografia
Alessio is a Postdoc at Boston University, where he works with Prof. Aldo Pacchiano. He has a Ph.D. in Electrical Engineering from KTH Royal Insitute of Technology (Stockholm), where he was supervised by Prof. Proutiere Alexandre. In his research, he studies problems of adaptive exploration in RL and Robust RL. Website: alessiorusso.net