What’s the game plan with AI? Limits and opportunities in AI
Artificial intelligence-powered computers can already beat the best mankind has to offer when it comes to chess and Go. But how did it get that way? And what can’t the best of AI do? Taking a look at how AI has developed its game-playing prowess can give us a few clues as to the next move.
Game theory with AI
Computer scientists have a long affair with games and have developed a number of programs pitting human intelligence versus the computer. But while making a program that excels at Tick-Tack-Toe is one thing, designing a computer program that can beat chess Grand Master is another. Given the intellectual intensity of chess, achieving this task has been the Holy Grail of programmers since the beginning of the computer era.
IBM takes the lead with Deep Blue
The first triumph for computers came when IBMs Deep Blue supercomputer out dueled Garry Kasparov in 1996. Not only did this take lots of computing power, it required LOTS of data – IBM engineers looked through over 700,000 grandmaster games for their needed info. In short, computer chess meant you were playing a compilation record of Greatest Hits of the Chess World. Of course, the computer did well – but it still needed someone to compile the great moves by chess champions of the past. Just think of this as the original Big Data approach – using a huge mass of data to reach a very specific conclusion.
Google goes for generic excellence at gameplay
DeepMind, Google’s AI-focused subsidiary has upset this equation with its latest rendition of AlphaZero, an algorithm that can achieve amazing game performance in chess, Go, and shogi (Japanese chess) – soundly beating world champions in each case.
Beating grandmasters was not the real achievement of AlphaZero. The most important features of this particular algorithm collection are that it is tabula rasa (that’s Latin for blank slate) and it is a generic AI. As a blank slate, AlphaZero starts out with no existing knowledge — except the specific game rules – and quickly learns as it goes. The concept of playing against itself over and over and learning from this experience is called “reinforcement learning” in computer terms.
Secondly, AlphaZero was designed to be a generic AI – easily used in a number of more specific processes due to its ability to automatically learn categories as it goes. That’s why its ability to excel at each of these three games without additional customization is such a big deal. AlphaZero can’t be applied to every task, but you can clearly get the idea that the software engineers want to come up with an AI that can be easily used for a wider range of activities.
Security as an AI-powered game with Avira
Avira AI is usually called Applied AI or machine learning – falling between the IBM and the Google examples. We use this for two primary uses – identify incoming threats and monitor individual smart device behavior.
Now in our third generation, Avira uses AI to analyze vast amounts of data, recognize patterns and anomalies, and provide users with a faster detection than is possible with traditional signature for antivirus. As with IBM and chess, we have a huge databank of malware samples which is used for machine learning our AI-powered detection engines.
Our new SafeThings security product uses AI to categorize smart devices, learn their data usage patterns, and detect anomalies. To top this off, the AI uses the information to automatically take the best action to secure the device without disturbing the owner. This is a machine learning type of AI where a less extensive quantity of data is needed.
Someone still has to make the rules
In all organized games – whether chess or malware detection – there are established rules. While some may be quite basic, others are more specific. For example, the knight in chess has a very specific move which is quite different from that of the king. Identifying malware has stringent requirements to prevent causing a false positive alert. While the AlphaZero needed only 24 hours of learning time for its exceptional performance in these three demanding games, it still started with the given set of rules.
Even the best AI cannot do what philosophers have termed creatio ex nihilo (Latin for creation out of nothing) — create or amend the rules for a more exciting time. Otherwise, we would be hearing about innovative AI—inspired ways to utilize your chessboard and player pieces.
While Google is showing us that AI can successfully be applied to a wider spectrum of our daily activities, it is a human that has the first – and last word – over the rules to the game.