Hierarchical Imitation Learning with Vector Quantized Models

Explore hierarchical imitation learning with vector quantized models in this 51-minute talk by Alexander Ilin from the Finnish Center for Artificial Intelligence (FCAI). Discover how intelligent agents can effectively solve complex tasks by planning actions on multiple levels of abstraction. Learn about a novel approach that uses reinforcement learning to identify subgoals in expert trajectories, associating reward magnitude with the predictability of low-level actions. Examine the vector-quantized generative model for subgoal-level planning and its application in complex, long-horizon decision-making problems. Understand how this algorithm outperforms state-of-the-art methods and can find better trajectories than those in the training set. Gain insights from Alexander Ilin, a Professor of Practice at Aalto University, whose research focuses on deep representation learning and model-based reinforcement learning.