Thursday, March 14, 2024

1120: Rationalism and Empiricism in AI...

 The battle between rationalism and empiricism has been ongoing in philosophy for centuries. Rationalism is the philosophical school of thought that assumes that only the use of human reason, ratio, can lead to sensible knowledge. Rationalism has no faith in the senses.


Empiricism is the view within the theory of knowledge according to which all knowledge ultimately comes from sensory experience, and conversely, that all scientific statements must be tested against experience.


By placing a strong emphasis on the development of the correct symbol system, early research into Artificial Intelligence fits well within the rationalist tradition.

  

The goal was not to allow an AI program to gain experience but to provide AI with the knowledge to successfully complete any experience.

   

In chess, the goal was to give AI a multitude of strategies to win the game, not to let AI play as many games of chess as possible in order to develop its own strategy.

    

With the emphasis on gaining experience, AI has developed over the past 30 years from rationalism to empiricism, which means, from knowledge that is given in advance to knowledge that is created through experience.

   

AI's victories in chess and Go are good examples of this. It was 1997 when Deep Blue, developed by IBM, won against Kasparov in chess.

  

In combination with special chess chips, which allow the program to look eight moves ahead, and search techniques to find the right move, victory over the mediocre-playing Kasparov is possible.

  

Due to the large amounts of pre-recorded knowledge put into it, Deep Blue is an exponent of rationalism.

   

Twenty years later, the company DeepMind, purchased by Google, came up with the program "AlphaGo". Go is a popular game in China, Japan, and South Korea

  

and is seen as a game that is more difficult for the computer than chess due to the many possible moves and game situations. Yet the program "AlphaGo" manages to beat the world champion.

   

The more successful version "AlphaGo Zero" goes one step further. The program is only given the rules of Go and is otherwise a "tabula rasa", a blank slate.

  

Expert knowledge is lacking, as are examples of matches played. "AlphaGo Zero" learns the game by playing against itself about 29 million times.

  

The program became an extremely successful player. So good, in fact, that Korean Go grandmaster Lee Sedol had to come to the conclusion in 2019 that "AlphaGo Zero" is unbeatable.

  

The empiricist approach to Artificial Intelligence turned the computer into an autonomous learning machine. And that raises some questions,

  

for example "Where does the computer gain its experiences?" or "How does the computer process and interpret its experiences?" or "Who is training the computer?". In other words, a completely new and barely explored area for Artificial Intelligence lies ahead of us.

   

Thank you for your attention again...