Google AlphaGo How It Works（AlphaGo如何工作）

　　Presenter: Shane (Seungwhan) Moon

　　PhD student

　　Language Technologies Institute, School of Computer Science Carnegie Mellon University

　　Useit知识库 报告地址：https://www.useit.com.cn/thread-11622-1-1.html

　　AlphaGo vs European Champion (Fan Hui 2--‐Dan)

　　AlphaGo vs World Champion (Lee Sedol 9-Dan)

　　Computer Go Al?

　　Computer Go Al - Definition

　　omputer Go Al - Definition

　　Computer Go Al — An Implementation Idea?

　　Computer Go Al — An Implementation Idea?

　　Process the simulation until the game ends, then report win / lose results

　　Choose the 'next action / stonew that has the most win-counts in the full-scale simulation

　　This is NOT possible； it is said the possible configurations of the board exceeds the number of atoms in the universe

　　Key: To Reduce Search Space

　　1. Reducing “action candidates” (Breadth Reduction)

　　1. Reducing “action candidates” (Breadth Reduction)

　　2. Position evaluation ahead of time (Depth Reduction)

　　F there is a function that can measure

　　Learning: P ( next action | current state )

　　(1) Imitating expert moves (supervised learning)

　　(1) Imitating expert moves (supervised learning)

　　(1) Imitating expert moves (supervised learning)

　　Looking ahead (w/ Monte Carlo Search Tree)

　　Use the networks trained for a certain task (with different loss objectives) for several other tasks

　　Lee Sedol 9-dan vs AlphaGo Energy Consumption

　　AlphaGo is estimated to be around ?5-dan

　　aking CPU / GPU resources to virtually infinity?

　　AlphaGo learns millions of Go games every day

　　What if AlphaGo learns Lee's game strategy