AlphaZero is the most high-profile AI development from DeepMind and is based on subtle algorithms of neural networks and artificial intelligence. The system is now considered the strongest player in chess and other board strategy games of a similar type. How can the game affect the development of innovation and new opportunities for technology? We will discuss this in this article.
Table of Contents
The Story of Alphazero’s AI Developments
DeepMind Technologies is a startup founded in 2010 by several London students. Founders Demis Hassabis, Shane Legg, and Mustafa Suleiman were passionate about developing artificial intelligence, so they experimented with combining machine learning, neuroscience, math, engineering, computing infrastructure, and modeling. This innovative approach attracted good investments; Elon Musk also noted the prospect of the project. In 2014, Google acquired DeepMind Technologies after Facebook halted negotiations back in 2013. And in the fall of 2016, the company transitioned to Alphabet and wiped Google from its name.
As a basis for developing their idea, the DeepMind team used computer chess games and variations of them, which they used to test out AI. One such program learned how to play more than 50 games, and Alpha Go was the first system that could beat a professional player – a live human – in a Go game. This program used a learning technique with a teacher and reinforcement to train it. The system exercised on human experience, learned to predict the player’s moves and then played against its own versions.
Progress has not stood still, and in 2017 the developers rolled out a new version of the program AlphaZero. It explains its name because the program was trained from scratch without using human experience. Developers trained the AlphaZero Neural Network to anticipate possible human choices and its own.
The DeepMind team claims that the AlphaZero version is by far the strongest player in Go and chess ever.
Neural Network Capabilities in Chess and Beyond
An artificial neural network is a way to recreate the simulation of the human brain using sophisticated algorithms. Interestingly, even the artificial NS is capable of self-learning. This process looks like an optimization task that uses analysis and clustering techniques from a maths perspective. An artificial neural network can analyze data using complex algorithms and perform operations on well-defined mathematical and fuzzy language systems.
The system decomposes the data into simple components; the algorithms form layers that analyze and transform the data. As learning progresses, some features become more complex and compact. Nevertheless, it turns out that the machine’s conclusions are similar to human conclusions. The AI development services can use these solutions to develop other areas unrelated to chess, Go, or games.
How AlphaZero Algorithms Work
AlphaZero learned to play chess very quickly. Initially, developers introduced only game rules into the system. They didn’t put any databases of games or libraries. After 24 hours, the program was already able to play with itself. We should note that at first, the moves were random. When one of the sides lost, the system calculated which moves were more effective in total.
At that time, AlphaZero’s main rival was Stockfish. It analyzed more than 70 million chess positions per second. Whereas the DeepMind product counted only good games, their quantity was less, but the quality was higher. As a result, the program spent half as much time processing as the opponent.
Unlike AlphaGo’s previous version, Zero’s version did not use data from games with humans or between human players. Instead, it skipped that step and started playing directly with itself. Soon the program surpassed the human level and defeated the champion version.
We can say that all the predecessors used human experience, while this version does not need it. It is no longer the heir to human achievement. AlphaZero’s victory is somewhat symbolic.
Also Read: QuickBooks Error 404
AlphaZero knew nothing about how humans played. Engineers only told it the rules and the goal. And after a short (measured in hours, not months) practice in “playing by itself” mode, it defeated programs that learned from humans.
Unlike previous DeepMind programs, AlphaZero was created as an algorithm capable of learning several task games at once, not just one. To do this, it was not trained to win but was given only basic knowledge of the game’s rules. AlphaZero then played with itself and developed its tactics.
Of course, the developers don’t fully disclose all the program’s secrets, but many note the high probability of using powerful search algorithms. When the neural network is in play, its metrics are constantly updated, calibrated, and then combined with the search algorithm.
AlphaZero is not limited to human knowledge alone but is capable of learning from scratch. Among the set parameters of the AlphaZero version were black and white chess/stones on the board only and a combined policy and value network to select moves and predict the likely winner after each move.
These steps helped optimize the system and improve its performance over its predecessors.
AlphaZero Algorithm Outlooks on Other Fields
Let’s talk about how the system can be trained from scratch without human experience. It means that we can apply it to any other domain besides chess. But, of course, the DeepMind team’s goal was not just to create a game-winner but to create an AI that would be effective and applicable as a unified program, roughly speaking, a “boxed solution.”
DeepMind’s creation didn’t just learn how to play like a human. Instead, it created its own approaches and strategies. As a result, AlphaZero learns a lot more in a short period than a human can.
Board games are not the only field where calculus and strategy are applied. There are areas far from entertainment that operate similarly. That is where the ability to learn from human strategy and experience and the machine inventing its effective method from scratch can come in handy. We can use similar algorithms in medicine, physics, and even art. AI is already trying to create paintings, music, and poems.
About the author:
Robyn McBride is a journalist, tech critic, author of articles about software, AI and design. She is interested in modern image processing, tech trends and digital technologies. Robyn also works as a proofreader at Computools.