For artificial intelligence (AI) systems to get to a point where they can solve problems better and faster than humans, first humans need to train the machines to use AI. One such way to train these machines is through continuous testing and reinforcement.
Simple games, such as Checkers, have been used to train AI systems through trial and error and as a result have made computers smarter, enabling them to get better at a number of things, not just the games the machines are trained on.
Through games, AI developers can test and compare the way computers and humans solve problems. By feeding the computer the game’s rules and teaching it the strategies humans would use, researchers can test how their AI systems match up to their human counterparts.
This is not a new way of testing AI systems - some of the first computers with AI learned to play games such as Checkers and showed researchers that computers with AI can do tasks just as well as humans, but faster.
The ability to complete tasks faster is down to the enhanced memory a computer has compared to a human.
From the beginning: teaching AI how to play games
In 1956, IBM’s Arthur Samuel, a Computer Engineer, taught a computer how to play Checkers ultimately showing that machines can learn simple tasks, and after practice, can do it better than humans. IBM’s research into computers was one of the first successful examples of machine learning.
Taking machine learning to the next level, IBM wanted to test its computers on a more complex game and set its sights on teaching a computer to play Chess. Testing computers on this game started in the 1950s, however, the technology giant struggled to make its computer system better than a human and it wasn't until 1997 that IBM’s supercomputer, Deep Blue, beat world chess champion, Garry Kasparov.
IBM’s famous software, Watson, was also tested and evaluated on its ability to play the television show Jeopardy!. IBM staffed a team of 15 and gave a three to five-year time frame to perfect this new software. This time was essential as the game proved hard for Watson because language was a difficult concept for computers to grasp. By 2010, Watson was successfully winning against Jeopardy! contestants.
It hasn’t just been IBM that has relied on different games to test its AI software, AIphabet’s AI developers at its DeepMind unit were determined to develop a computer that was successful at playing Go, a strategic board game more challenging than chess. DeepMind taught a computer programme the moves champions used and had it practice playing against humans.
In 2016, the programme came up with new moves humans hadn’t thought of and eventually beat 18-time world champion Lee Sedol. Four years later, after extensive research into reinforcement learning, DeepMind were able to develop a programme that could play Go, Chess, Shogi and 57 games by Atari - without being taught a single rule of the games.
AI and gaming: the importance of reinforcement learning
Reinforcement learning is key as companies look to use AI to solve different business problems. This learning works as a computer keeps playing a game, through trial and error, it can learn on its own and find ways to play a game not thought of by humans.
The reinforcement comes in the form of feedback from the researcher that tells the computer how it is doing. To become better than humans, computers keep playing a game over and over until it can’t beat its own record.
Over time, the computers can learn a winning strategy than even the top humans are yet to think of. Despite this, reinforcement learning is no substitute for human intelligence as it cannot replicate emotions, creativity and common sense.
Researchers need to take this into consideration as they develop new AI systems and ensure that this type of learning doesn’t negatively impact the computer. Together, human intelligence and AI-enabled computers, can come up with a number of solutions to combat a variety of problems, and when applied in a business environment can lead to a number of benefits.