Explore a groundbreaking approach to AI planning in this 26-minute video explanation of the paper "Divide-and-Conquer Monte Carlo Tree Search For Goal-Directed Planning." Delve into a novel generalization of Monte Carlo Tree Search (MCTS) that revolutionizes problem-solving by recursively dividing complex tasks into manageable sub-problems. Learn how this method deviates from traditional step-by-step planning, instead focusing on identifying optimal intermediate goals. Discover the algorithm's unique ability to improve imperfect goal-directed policies through strategic sub-goal sequencing. Examine the concept of Divide-and-Conquer MCTS (DC-MCTS) and its application in both grid-world navigation and challenging continuous control environments. Gain insights into the flexibility of planning strategies and their potential to outperform sequential planning approaches.
Overview
Syllabus
Intro
What is planning
The algorithm
Finding the next action
Building your search tree
Search over subproblems
Subdivide
The Catch
Deep Learning
Training
Taught by
Yannic Kilcher