Title: Distributional reinforcement learning: A richer model of agent-environment interactions

Title: Distributional reinforcement learning: A richer model of agent-environment interactions
Abstract: Biological and artificial agents alike benefit from treating their environment as a stochastic system. In reinforcement learning, we instantiate this principle by modelling the environment dynamics and total reward (the return) as random quantities. Where the classical treatment focuses almost exclusively on the expected return, a much richer picture emerges when we instead consider the entire distribution of returns. I will give a technical overview of the computational concerns and solutions that arise when we design agents that learn return distributions. Following this, I will review recent experimental results, from robotics to computational neuroscience, illustrating the broad benefits of studying and designing agents under the distributional lens.

Bio: Marc G. Bellemare leads the reinforcement learning (RL) group at Google Research in Montreal, Canada. He is adjunct professor at McGill University and Université de Montréal, a core industry member at the Montreal Institute for Learning Algorithms (Mila), CIFAR Learning in Machines & Brains Fellow, and holds a Canada-CIFAR AI Chair. At its core, his group studies how artificial agents can be designed to operate in complex, time-evolving environments. Marc received his Ph.D. from the University of Alberta, where he developed the highly-successful Arcade Learning Environment benchmark. During his subsequent tenure at DeepMind in London, UK he made a number of pioneering developments in deep reinforcement learning, in particular proposing the distributional perspective as a richer model of agent-environment interactions.

свернуть