Milestones in Machine Learning
With the advent of higher computing power, machines are now able to perform tasks that were believed to be possible only by humans. Machines are now capable of beating the best humans in Chess and GO, games that were once thought impossible for computers to master. Below I highlight a few watershed moments in Machine Learning History.
Deep Blue
The Deep Blue was a chess playing computer developed by IBM. Development began in 1985 under the name Deep Thought and in 1989 it was renamed Deep Blue. In 1996 it made history after defeating Garry Kasparov, the world’s best human chess player, and, therefore, becoming the first computer to win both a chess game and a chess match against a reining world champion under regular controls.
Deep Blue used custom Very Large-scale Integration (VLSI) chips and an alpha-beta search algorithm to find its best moves. This allowed the computer to play strategically and more human like, something that at to that point computers where not yet capable of doing. In fact, the computer played so much like a real human that Kasparov, after losing, accused the Deep Blue team of cheating, saying that a grandmaster human chess player had to be behind the play that led to his defeat.
Watson
After defeating the world’s best chest player, IBM went searching for another challenge. In 2006 they began the development of a supercomputer that could answer questions in natural language, a task that at the time seemed impossible for computers to achieve because of the double meaning, and wordplay usually found in language. The supercomputer, named Watson after IBM’s founder Thomas J. Watson, went on in 2011 to win in first place in the US quiz show Jeopardy!, the first time a computer beats Jeopardy! champions Brad Rutter and Ken Jennings, who, with 74 consecutive wins, holds the record for the longest winning streak on Jeopardy!.
Watson was able to achieve this feat using more than 100 algorithms that search for possible answers and then ranks the results by how likely they are to answer the question. It is also capable of cross referencing against time or space, so if someone was not born in a relevant year, then it is less likely to be correct.
AlphaGo
After computers were able to beat humans in chess, the next challenge was to defeat humans at the board game of Go, which is an abstract strategy game invented in China and is more complex than chess in terms of the numbers of possible moves — 10^123 in chess vs 10^360 in Go. This level of complexity was too difficult for Traditional methods such as the one used for Deep blue, Alpha beta pruning. AlphaGo used a Monte Carlo Tree Search algorithm, which is improved by artificial neural networks. In 2016 AlphaGo made history after defeating, for the first time, a professional human player on a full-sized board without handicap (which is giving in the form of stones and compensation points to offset the strength difference between players of different ranks) in a match against European champion Fan Hui. Then, in 2016, it went on to defeat the world champion from South Korean Lee Sedol. After his defeat, Lee Sedol was quoted refereeing to AI as “an entity that cannot be defeated.”
Beyond AlphaGo
After AlphaGO’s success, the DeepMind team crated a successor to AlphaGO called AlphaZero, which, in 2017, was able to defeat AlphaGo 100 games to 0 just after a year of Lee Sedol’s defeat by AlphaGO. AlphaZero can generalize and master other games besides Go, including chess, shogi (also known as Japanese Chess). Alpha Zero did not learn from human players as did AlphaGO, and instead it started with the game’s rules and learned by playing against itself.
More recently DeepMind developed Muzero, which can master games without knowing their rules. It has matched AlphaZero’s performance in chess and shogi, and it has improved in Go. It was also capable of mastering all 57 original Atari games.