Artificial intelligence (AI) systems that play video games have reached a new phase. For example, the traditional board game Stratego, which is more complex than chess and Go and more agile than poker, has now been mastered. The researchers introduce DeepNash, an artificial intelligence agent that learned the game to the level of a human expert by playing against itself.

Model-free deep reinforcement learning and game theory are the foundation of DeepNash's innovative methodology. Moreover, its play style converges to a Nash equilibrium, making it very difficult for an adversary to take advantage of it. DeepNash has worked so hard that she is currently ranked among the top three human experts on Gravon, the largest online Stratego platform in the world.

Board games have historically served as an indicator of advancements in artificial intelligence (AI), allowing us to examine how humans and machines create and execute plans in a controlled environment. However, Stratego, unlike chess and Go, is a game of incomplete information, as players cannot directly examine the identities of their opponent's pieces.

This difficulty has prevented other AI-based Stratego systems from surpassing the amateur level. It also means that a very successful artificial intelligence technique known as "game tree search," which has been used to master many games with perfect knowledge, is insufficiently scalable for Stratego. DeepNash goes well beyond game tree searching for this reason.

Mastering Stratego has benefits beyond the gaming realm. To achieve their purpose of solving intelligence to advance science and serve society, we must develop advanced AI systems that can work in complicated, real-world scenarios with minimal knowledge of other agents and individuals. Their study demonstrates how DeepNash can be utilized in tentative plans to balance outcomes and solve complex challenges effectively.

Conclusion

We can immediately extend the researchers' groundbreaking R-NaD strategy to other two-player zero-sum games of either perfect or imperfect knowledge. However, DeepNash was designed explicitly for the clearly defined realm of Stratego. Large-scale real-world problems, which are frequently characterized by incomplete information and astronomical state spaces, can be addressed by R-NaD by generalizing much beyond two-player game environments. They also hope that R-NaD will help open up new AI applications in fields. 

Furthermore, the researchers want to better integrate AI's problem-solving abilities into their intrinsically unpredictable reality by developing a generalizable AI system that is resilient in the face of uncertainty.