In the game of Nim, there are specific optimal moves for any given board setup. If you don’t take one, you hand control to your opponent. They can win if they stick to their best moves. This strategy hinges on a mathematical concept called the parity function.
Interestingly, research by Zhou and Riis suggests that the usual training methods for complex games like chess don’t translate well to Nim. They found that AI trained on a five-row Nim board improved quickly. But when they added just one more row, the AI’s progress slowed significantly. Once they reached a seven-row board, the AI’s performance plateaued after 500 self-play games.
To illustrate this issue, the researchers replaced the system that suggested moves with a random one. On the seven-row board, both the trained AI and the random AI performed almost the same after 500 games. This indicated that the AI struggled to learn effectively once the game board became too complex. Initially, in a seven-row setup, three potential moves all led to a win. However, the AI assessed each of these moves as nearly equal in value, showing its inability to differentiate them.
The key takeaway is that mastering Nim requires an understanding of the parity function, which the typical training setups for chess and Go don’t teach effectively.
This problem isn’t exclusive to Nim. Zhou and Riis also spotted similar issues in chess AIs trained in the same way. They found instances of “bad” moves—like missing a checkmate or making poor endgame decisions—that the AI initially valued highly. It was only when the software analyzed several moves ahead that it caught those mistakes.
This raises questions about how AI learns different games. The complexity of a game’s rules can significantly affect an AI’s ability to improve. For instance, experts in AI and gaming are now rethinking training methods to better equip AIs for various games with complex strategies.
As AI continues to evolve, understanding these learning limitations is crucial. After all, improving AI learning can lead to smarter systems in other applications, from healthcare to finance. Keeping an eye on these developments is essential for future advancements in artificial intelligence.
For more on the challenges faced in training AI for complex games, you can read about it in this research paper.
