Shulou(Shulou.com)11/24 Report--

CTOnews.com, December 4 (Xinhua)-- artificial intelligence "DeepNash", developed by the British company DeepMind, has defeated expert-level real-life players in "Army Chess / Strategic go (Stratego)" through "bluff" fraudulent means. It can even deliberately sacrifice important chess pieces in order to win, which is of great significance in AI research.

The relevant research report has been published in the international scientific journal New Scientist (New Scientist), DOI:10.1126/science.add4679.

The scholars of the study say it represents an "extraordinary result" because people in Stratego circles believe that "this cannot be achieved with modern technology".

In Stratego, each player controls 40 different levels of chess pieces to play on a 10 × 10 square board. If the two sides encounter on the chessboard, the higher-ranked pieces will beat the lower-ranked pieces, but you can't see the identity of each other's pieces, but this can cause countless "accidents", which cannot be easily cracked by exhaustion.

To put it simply, Stratego is a strategic board game similar to the military flag. its rules are simple enough for young children to play, but the depth of its strategy is difficult for adults to understand. From a "strategic" point of view, Stratego has more variables and more difficulty than chess, go or playing cards.

"the most surprising behavior for us is that [AI] is able to sacrifice valuable pieces to get information about the strategies made by opponents," said Julien Perolat, a DeepMind researcher.

For many years, Stratego has been one of the next frontier fields of AI research, which aims to test whether AI can make relatively radical, cautious or long-term behavior like human beings. in the final analysis, it is to examine the ability of AI to judge game logic and make on-the-spot decisions.

Here, Julien Perolat and his colleagues introduce a new method that allows AI to learn the game better. Using this new method, they developed an AI robot called DeepNash, which can demonstrate human expert performance in the most complex version of the game, "Stratego Classic".

CTOnews.com learned that the core of DeepNash is a reinforcement learning algorithm "R-NaD". In order to make DeepNash,Perolat and the team will combine R-NaD with deep neural network architecture, so as to make it learn and think about game strategies at the high-end competitive level.

In the end, DeepNash conducted game tests with a variety of state-of-the-art Stratego robots and professional players, and it eventually beat all the other robots and beat all expert live players on Gravon (an Internet gaming platform and Stratego's largest online platform). As of the time of the report, it was in the top three on the platform, while its opponents had no idea that there was an AI across the network.

In addition to that, DeepMind AI has also won 97% against top strategy game robots, including some that have won world championships.

"good players tend to remember their opponents' pieces and predict their core layout," said Georgios Yannakakis of the University of Malta. "DeepNash does both very well-of course, AI may have an advantage over people in memory, and he will play games in interesting and unpredictable ways, such as bluffing."

Tuomas Sandholm of Carnegie Mellon University in Pennsylvania says DeepNash game theory can prove that AI is also useful in non-game situations with other people, such as business and defense.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.