01-31-2021, 01:26 PM
Quote:Would theorems (of positional properties) make useful features for Polygames?
In principle, I'd be inclined to say yes. If you can obtain reliable, accurate additional information about a game state, which can be relevant for the decision-making process of a game-playing agent, then providing that information as additional inputs to a neural network would likely allow it to learn more rapidly than if it had to automatically discover that knowledge through training with more raw inputs.
The main issue I see is that I would expect automated theorem proving based on Ludii's game descriptions to be rather difficult, much more so than with a game description language that is already logic-based like GDL. As Cameron already mentioned in the post above, the semantics of every single token that can be used in a game description file for Ludii are defined just by the corresponding Java code; the semantics are not defined in terms of logic-based propositions. I suppose that somewhere deep down every piece of Java could would have clear, strict semantics... but this is not readily and explicitly available in a logic-based format, and I imagine it would be extremely complicated to get it into such a form. Maybe I'm missing something though, this is not my area of expertise.