You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
So if you have played around with the algorithm parameters and let it run for an extended period of time you will realize that the evolution stagnates at a certain point (coinciding with the screen completely filling) and fitness increases extremely slower from that point on. This is due to the leap from simply filling the screen to actually clearing lines being way too high and requires a massive amount of new innovations to occur. Rather than throwing the entire project away, I'm thinking of trying an algorithm to let it learn the required innovations in what I'm coining as "Supervised Neuroevolution".
Essentially instead of dumping the entire Tetris state with all 20 rows and 7 pieces, we start out with 3-4 rows and 1 piece, slowly introducing new pieces and rows as the overall fitness improve to a point deemed worthy of further evolution. The input neurons will still take the same index but will be considered disabled and not be allowed to be picked for link mutations.
Hopefully this works out or we'll have an exciting report and a non-existent demonstration.
The text was updated successfully, but these errors were encountered:
So if you have played around with the algorithm parameters and let it run for an extended period of time you will realize that the evolution stagnates at a certain point (coinciding with the screen completely filling) and fitness increases extremely slower from that point on. This is due to the leap from simply filling the screen to actually clearing lines being way too high and requires a massive amount of new innovations to occur. Rather than throwing the entire project away, I'm thinking of trying an algorithm to let it learn the required innovations in what I'm coining as "Supervised Neuroevolution".
Essentially instead of dumping the entire Tetris state with all 20 rows and 7 pieces, we start out with 3-4 rows and 1 piece, slowly introducing new pieces and rows as the overall fitness improve to a point deemed worthy of further evolution. The input neurons will still take the same index but will be considered disabled and not be allowed to be picked for link mutations.
Hopefully this works out or we'll have an exciting report and a non-existent demonstration.
The text was updated successfully, but these errors were encountered: