AI - A story of four games

AI is not new; it has a 60-year history. It has seen many ups and downs.

At the middle of last century (1950), the world came out of the world wars. There were two legacies of the wars - the nuclear bomb and the computer. Both changed the world dramatically.

When scientists invented computers, AI was the first application on their mind. A single person is often credited for fathering both computers and AI - Alan Turing. His 1950 paper ‘Computing Machinery and Intelligence’ pioneers AI thought and proposed the famous Turing Test.

 In 1956, a conference at Dartmouth College (near Boston in US) declared the name Artificial Intelligence officially. In these early days, there was great enthusiasm in scientists and supporters about AI. They were super-confident that AI will reach human capabilities in a decade. How innocent that sounds now!

Initially, AI looked like it is living up to the promise. Scientists wrote programs that could play the game of checkers or prove mathematical formulae. The goal of matching human intelligence seemed near. 

But by 1975 the enthusiasm died. Supporters realised that the job is much tougher than it seems. The ground reality was that the ecosystem was just not ready. Hardware was not powerful enough, techniques were limited. The early promises were not fulfilled. AI looked like a failure. A period of neglect – no funding and no jobs- began. This is called the 'first AI winter'.

In the eighties, AI solutions called 'Expert Systems' became popular. These brought in the first billion dollar revenues for AI. It became an industry. Expert Systems were used everywhere - medicine, telecommunications, IT. A language called LISP formed the foundation of these systems.

In the early nineties, the second bust came. Expert systems proved to be too simple for the increased complexity. Neural Networks, which showed promise, suddenly fell out of favour. The Second AI winter started. All seemed to be lost.

AI was saved by an ancient human pastime – games. The love affair of AI with games had started earlier, but in the fading years of the last millennium, it became a really strong relationship. The story of four games began in 1997, when an AI powered machine Deep Blue by IBM defeated the then chess champion Garry Kasparov. This was the first victory of AI on a reigning human champion. The event was widely publicized. It gave the people an inkling that AI may be something real and practical.

As the new millennium dawned, machines became more powerful, data volumes grew and internet made it easy to share data. Scientists started using different techniques generally called Machine Learning rather than the Goof Old Fashioned AI (GOFAI) methods. All this led to a change of fortunes, and the harbinger of the new age was another game.

It was IBM again, who with their monster-computer Watson won a television game called Jeopardy! in 2011. This is a kind of reverse quiz show. Example:

Jeopardy host: "Donald Trump"

Participant: "Who is the current American President?"

Jeopardy host: "Canis familiaris"

Participant: "What is the scientific name of common dog?"

When it was initially made, Watson was a 'Knowledge Based System' (KBS), as it is called in AI parlance. It had the ability to acquire large amount of knowledge from various sources, including the internet. It could then utilise this knowledge for a variety of purposes, such as answering queries. Now Watson is IBM's flagship AI product that incorporates many more capabilities than Q&A.

Since 2011, AI started enjoying sunshine again. This time, it reached both the consumers and enterprises alike. A slew of startups emerged, furiously working of research as well as applications. The funding for the startups took a jump, which again co-insides, or maybe even caused by a game company.

In 2014, Google acquired a company called DeepMind reportedly for 500 million dollars. DeepMind used a variety of Machine Learning called Reinforcement Learning to learn games by watching them being played. Google named the system AlphaGo and continued the research.

Last year, in May 2017, AlphaGo won a Chinese game called Go that is considered very hard to master even for humans. This win was important from another point of view. While both Deep Blue and Watson used rule based programming, AlphaGo learned from experience, using Reinforcement Learning (RI). The win thus clearly marked a shift of paradigm in computing.

The fourth game, played in December 2017, ushers yet another paradigm. It was AlphaGo again this time that beat a champion in chess, in a throwback to 1997. But this time it was not a human champion (human no longer play against AI now), but another chess program, the best chess playing program in the world called Stockfish 8. What was special was how it learned to play.

AlphaGo practiced against various chess programs and mastered the game in a flat four hours. In other words, it became good enough to beat the champion from zero in four hours! This is a demonstration of the new super powerful capabilities that AI systems can acquire through learning.

For the last twenty years, games have been the flagship events for AI. I think the attention will now shift to robotics. There might be a day, in not far away future, when the FIFA World Cup champions will play against a team of robot players. We can only guess the result.