Tractable Novelty Exploration Over Continuous and Discrete Sequential Decision Problems
Explore the latest advances in width-based planning algorithms for sequential decision problems in this 55-minute lecture by Dr. Nir Lipovetzky, Senior Lecturer at the University of Melbourne's School of Computing and Information Systems. Delve into the world of AI planning, focusing on structural exploration features rather than goal-oriented heuristics or gradients. Learn about state novelty evaluation and its exponential nature, and discover two key advancements: defining state features for continuous dynamics and developing polynomial approximations of novelty through sampling and bloom filters. Compare the performance of polynomial planners in discrete sequential decision problems with state-of-the-art deep reinforcement learning algorithms using OpenAI Gym benchmarks. Gain insights into how width-based planners can achieve comparable policy quality with significantly reduced computational resources.