Overview
Save Big on Coursera Plus. 7,000+ courses at $160 off. Limited Time Only!
Explore cutting-edge approaches to solving high-dimensional optimization problems in this 43-minute conference talk. Delve into recent advancements that combine deep neural networks with reinforcement learning and search methods to outperform traditional human-driven solutions. Discover how these innovative techniques are applied to various domains, including online job scheduling, neural architecture search, and black-box optimization. Learn about the Latent Space Monte-Carlo Tree Search (LaMCTS) algorithm, its simple usage, and performance in optimizing linear policies for Mujoco tasks. Gain insights into the challenges of learning action spaces, multi-objective optimization, and the limitations of current approaches. This talk provides a comprehensive overview of state-of-the-art methods for tackling complex optimization problems in design, operations research, and scientific exploration.
Syllabus
Intro
Reinforcement Learning
GOOGLE TEACHES AI TO PLAY THE GAME OF CHIP DESIGN
Black box Optimization
Properties of Black box functions
Latent Space Monte-Carlo Tree Search (LaMCTS)
Simple Usage
Another example
Motivating Examples
The Meaning of Learning Action Space
How to learn the action space?
Different Partition → Different Value Distribution
Learn action space
Approach
Sample in a Leaf
Performance
Optimizing linear policy for Mujoco tasks
Limitations
Multi-Objective Optimization
Taught by
Open Data Science