So it turns out that my definition of PacMan is not relational. I sort of always knew this, but now it is explicit. The representation begin used here is Feature-Based. Anyway, the point of this post is that if I were to migrate it to relational, I doubt I would be able to use cross-entropy (without changing the cross-entropy algorithm explicitly).
Actually, scratch that. I would still be able to use cross-entropy, but the generators would be enormous. Because, the generator would have to store every relation, in every form, resulting in massive generators and slow convergence.
However, I have realised that I need to use relational learning if I am to follow my initial proposal – using Cyc. By using Cyc, I could utilise typed logic [location(X) would only work for physical X]. So I think first, i should attempt to change PacMan into a relational setting, and see what I get. Following this, I could attempt to use Cyc, although using simple object inheritance may even work.
I have a feeling that this evolutionary path seems to be working, and I could even boost it by using other techniques during the process.