Artificial Intelligence Can Become Even More Perfect
After the success of Libratus (a poker bot that effectively solved the task of defeating a man in Texas Hold’em) artificial intelligence can become even more perfect.
This will happen if an international group of scientists in the field of computer science copes with the task of deciphering the team game scheme. Having witnessed the triumph of the poker machine created at Carnegie Mellon University last January, Jacob Crandall and his team want to thoroughly study all aspects of the team game.
In an interview with New Atlas, Crandall explained that artificial intelligence (AI) must learn to work in interaction with people to solve a problem, and not just defeat them, as Libratus did.
Using a new algorithm called S#, Crandall and his partners analyze the role played by interaction and compromise in the prisoner’s dilemma. In the framework of the experiments, the following connections were tested:
In such a manner the researchers wanted to evaluate the relationship during the game.
One of the key indicators in the prisoner’s dilemma is the decision of the player to make a move in their own interests or in the interests of the team. When there is a threat that a move in their own interests can aggravate the position of the player, a dilemma arises: is it worth playing as a team to achieve a better overall result or not.
The researchers’ ultimate goal is to understand the mechanism of interacting AI with people and find out what artificial intelligence needs to get the communicative skills.
Tit for Tat
S# was forbidden to lie for the sake of experiments. The algorithm included a series of phrases known as “cheap words”. When the computer realized that his partner was cooperative, he gave a positive reaction. However, any unfair act of the human player caused contempt in the computer which it expressed with phrases like “You will pay for it.”
In such games, the machines learn the importance of moral principles in teamwork, as well as the benefits that such work provides. When people played with S# without knowing about it, they began to interact with the partner more, hearing a positive reaction to their actions.
Crandall and his team hope to better understand the relationship between people during the game. Then, using this data, they plan to improve the AI’s communication with people.
Libratus has shown that computers are now capable of defeating people in zero-sum games such as poker. S# is designed to improve AI to a level where it will be able to work with people, rather than compete with them. The main goal is to create such artificial intelligence programs that would not only be smarter than us but also help us achieve the best result in solving any of our problems.