AlphaZero and Machine Learning

Tony Su


March 14, 2019

What is Machine Learning?

My previous nEXT presentation on Data Analysis, 2015

What is NOT Machine Learning?

Relational Databases

Online Analytical Processing (OLAP)

Data Mining (Splunk, Microsoft BI )

NoSQL Databases (Cassandra, MongoDB)

Search (Yahoo, Google Internet Search)

Big Data Analysis (Hadoop, Elastic)

Cloud Storage


Conventional ML

Typically requires a Data Scientist


AlphaBeta computations, including MinMax



Enormous number of computations


AlphaGo ML

Typically requires a Data Scientist

Supervised Training

Convolutional Artificial Neural Networks

Monte Carlo Tree Search (MCTS)

   1987 Bruce Abramson

    Ahead of its time, computers not powerful enough

Enhances Exploration at cost of some Exploitation

Particularly well suited for GPU computing

First version implementing Deepmind concepts

AlphaZero ML

Typically requires a Data Scientist

Unsupervised Training(debatable)

Convolutional Artificial Neural Networks

A Policy Network and a Value Network

Introduces concept of training composed solely of self-play.

Devoid of human guidance

Based on adversarial training, given only basic rules.

The algorithms and NNs are completely self-developed by the machine

AlphaZero History

AlphaGo Zero

2015 Created AlphaGo
2016 Beat a 9th dan human Go player
2017 Beat World Go Champion in a match
2017 Created AlphaGo Zero
2017 AplphaGo Zero beat AlphaGo 100 games to 0 with less computing power
2017 AlphaZero (modified AlphaGo Zero) plays chess and shogi

Google AlphaZero Use

  • Wavenet - Modeling raw audio waveforms​
  • Before AlphaGo, was too computationally intensive to produce good results
  • Google Assistant (speech recognition)
  • Data Center cooling efficiency
  • Google Play User recommendations
  • Future Android features Adaptive Battery and Adaptive Brightness
  • Deepmind Health
  • 2016-current analyzing patient data for various purposes including diagnosis and prognosis, patient management

Potential AlphaZero Use

Example of recent successes in Computer Vision
voice recognition
facial recognition​

Potentially any use that can be represented as a game

Self-driving vehicles

War games


The Real World Proof


Current Head to Head Stockfish vs Lc0

Top Computer Chess Engine Championship (TCEC)

Items to Note

Number of nodes (positions evaluated)

Speed - Number of nodes evaluated per second



Open Source



openSUSE Linux


Installation Summary basic steps (all platforms)

  • Download or Compile Binary
  • Download a Weights file
  • Place the Weights file in the same folder as your Binary

Option 1

Run commands  directly on the Binary

Option 2

Connect to a Chessboard frontend

Ethics and the Future

China and the USA are in the lead
60 minutes report on China's efforts resulting from one entrepreneur's work
China study identifying cancerous nodes similar to highly trained doctor (97%)
Number of patents per year.
Ethical concerns, eg the value of Privacy and how it's regulated

Mobile Hardware Platforms

Snapdragon 855 specs and descriptions

Google Pixel 3 Visual Core

Snapdragon 855 vs Apple  A12 Bionic



A brief history of Data Analysis

Google and Deep Mind

Google's Machine Learning site

MIT Sloan Analytics Conference 2019

60 minutes AI and China

Computer Vision and Cancer detection


Q & A