Adaptive Bayesian Intelligence Team Seminar(Talk by Henrique Donancio, INRIA Grenoble Alpes).

【Team】Adaptive Bayesian Intelligence Team
【Date】2025/August/15(Friday) 17:30-18:30 (JST)
【Speaker】Talk by Henrique Donancio, INRIA Grenoble Alpes

Title: Deep Reinforcement Learning

Abstract:
Deep Reinforcement Learning (RL) has achieved remarkable success in simulated domains but
still faces major challenges in real-world applications, including safety constraints and sample
inefficiency. In this talk, I present the Pump Scheduling Problem, a real-world testbed derived
from an operational water distribution system, designed to evaluate RL methods on aspects
often overlooked in synthetic benchmarks such as Atari and MuJoCo. I then introduce Dynamic
Learning Rate for Reinforcement Learning (LRRL), a meta-learning approach that leverages
adversarial bandits to adapt learning rates on the fly, improving training efficiency. Finally, I
discuss advances in distributional RL, highlighting Gaussian mixture models for return
representation and novel probability metrics. Together, these contributions outline a path toward
more reliable, adaptive, and data-efficient RL.