Photo by Pixabay on Pexels

Researchers are exploring a novel approach to artificial intelligence training: using randomly generated games. Inspired by games like Chesh, created by Damien Sommer, this method aims to cultivate adaptable AI capable of mastering unfamiliar environments. Chesh, a game that procedurally generates chess-like variations with different board sizes and piece mechanics, presents a complex challenge for AI. The goal is to force the AI to develop meta-strategies applicable to a wide range of scenarios, rather than memorizing specific moves. This approach mirrors the training of systems like AlphaGo Zero, which learned to play Go from scratch through self-play. Although Chesh is no longer available, the concept highlights a promising path for developing more robust and versatile game-playing AI, prompting developers to create similar environments for AI training. The original idea was discussed on Reddit: https://old.reddit.com/r/artificial/comments/1nytvs9/training_ai_on_randomly_generated_chess/. More information about Chesh can be found at https://www.damiansgames.com/#/chesh/ and AlphaGo Zero at https://deepmind.google/discover/blog/alphago-zero-starting-from-scratch/.