Sounds like a question that would spark an intense debate. While highly subjective, there might be be a way to evaluate what could make some games played “better” than others. All you need to is ask NM Ionatan Giurgiu, who has developed a program that considers the following:
First, I need to run an engine analysis on the game to see how the evaluation changes. Then, run a “difficulty” scorer that uses Lc0 and Maia3 neural networks to estimate how difficult each move is. This information is used to get a “baseline score”, where a perfect game that’s easy might get around 10 points. And a perfect game that is incredibly difficult, around 20 points. Mistakes made in difficult positions are forgiven more than mistakes in easy positions.
That baseline score is also multiplied by a number, derived from various possible game features, including:
- King hunt
- Abnormality
- Sacrifices
- Tension
Also, bonuses can be received if the game includes:
- Brilliant moves (only-moves with a decent sacrifice score)
- Underpromotions that are only-moves
- Quadrupled pawns or a pawn chain of length 6.
- Castling after move 40.
- En passant mate
The feature scores and bonuses are added together and dampened using a logarithm, so that high feature scores don’t completely blow up the score.
This way, a badly played game that was very interesting can score as high, or even higher than a perfectly played game that was easy, with no interesting features.
