Video game-playing artificial intelligence (AI) systems have advanced to a new stage. For instance, it is now possible to learn the age-old board game Stratego, which is more challenging than chess, Go, and poker. The researchers describe DeepNash, an artificial intelligence agent that by competing against itself, learnt the game to the level of a human expert.
The novel approach of DeepNash is based on model-free deep reinforcement learning and game theory. Furthermore, it is highly challenging for an adversary to exploit it because of the way it plays, which converges to a Nash equilibrium. DeepNash has put in so much effort that, according to Gravon, the world’s largest online Stratego platform, she is currently ranked among the top three human experts.
Since they allow us to compare how people and machines make and carry out plans in a controlled environment, board games have historically functioned as a leading indicator of developments in artificial intelligence (AI). However, because players are unable to directly inspect the identities of their opponent’s pieces, Stratego is a game of imperfect information, unlike chess and Go.
Other AI-based Stratego systems have been unable to go past the amateur level due to this issue. Additionally, it implies that a highly effective artificial intelligence method called “game tree search,” which has been used to completely master a variety of games, is not adequately scalable for Stratego. For this reason, DeepNash extends well beyond game tree searching.
Beyond gaming, mastering Stratego offers advantages. We must create sophisticated AI systems that can operate in challenging real-world situations with little prior knowledge of other agents or people in order for them to fulfil their goal of solving intelligence in order to advance research and help society. Their research indicates how DeepNash may be efficiently used in tentative plans to balance outcomes and address difficult problems.
Conclusion
We extend the pioneering R-NaD technique of the researchers to other two-player zero-sum games of either perfect or imperfect knowledge. DeepNash, on the other hand, was created specifically for the well-defined Stratego world. R-NaD can be applied to large-scale real-world problems that are frequently marked by incomplete information and astronomical state spaces by expanding its application beyond two-player game environments. Additionally, they anticipate that R-NaD will contribute to the development of new AI applications.
Furthermore, by creating a generalizable AI system that is resilient in the face of uncertainty, the researchers hope to better mesh AI’s problem-solving capabilities with their inherently unpredictable reality.
