Here is a quick list from the perspective of LLMs playing chess:
- PGN provides game history, context, and metadata, aiding in strategic learning.
- FEN only describes a single board position, lacking historical context.
- The development of a game using FEN would be a list of fully specified chess positions, over the whole chessboard, with only few changes between the corresponding FEN strings.
- PGN using Algebraic Notation (AN) represents gave development move-by-move.
- I.e. much easier for sequential prediction.
- PGN allows LLMs to learn from move sequences and game annotations easier than with FEN.
- (Related to the point above.)
- PGN is versatile for game simulation; FEN for specific position analysis.
- PGN integrates well with extensive game databases for LLM training. It seems to me that:
- Many PGN databases exist, which are rich sources for LLMs to learn from a wide array of games, providing a vast corpus for training or analysis.
- While FEN is used in chess engines and tools for position analysis, its use is more limited in scope. (Hence, WWW representation.)
BTW, "Wolfram/Chess" is using (strict) FEN strings as input. Looking at the documentation of Chessboard no other type of input is allowed. Ideally, other formats would be allowed soon...
