The neural nets have nothing to do with it. The idea is not to play any complete number sets that have already been played.
To pick the individual digits, for say Mega Millions....
1) Use only Mega Millions draws since 10/22/2013 as that is the date that the game changed to 75 numbers and 15 mega or bonus numbers.
2) Re-arrange my draw history from lowest digit to highest digit for the first 5 digits, if not already done so by the state or web site that I am getting my draw history from.
3) Count how many times each digit showed up in the first position, keeping only the digits that showed up with the highest count. Do this action for each of the 5 number positions, try to make sure no numbers repeat from one position to the other. For example if the number 23 shows up in position 2 and 3, keep 23 in the position with the highest count, eliminate the other. If both have the same count use the lowest position with the most recent appearance of 23 in this example.
4) Do a skip and hit chart for the numbers in each position, keeping the digits with the lowest skips in the last 30 draws.
5) Eliminate numbers in the last/latest draw.
6) Get the sum numbers for all of the prior draws (note step1), count how many times each one showed up keeping the sums with the highest count. An alternative to this idea is to use Gail Howards most probable sums which would be 141-240 for a 5/75 game according to her web site.
7) Make a hot and cold list of the last 30 games (75/5*2) for each position.
8) Construct each number you want to play with a mix of 2 hot 3 cold or 2 cold 3 hot using 2 odd 3 even or 3 odd 2 even for each number set making sure they add up to the sums in step 6, using the numbers you have left from the prior eliminations.
Is this idea any better than a quick pick? Don't know, have not made a spreadsheet for it. It may create too many combinations to be practical. Some one here at LP may have already made a workout for this or something similar.
Its just an idea, maybe it could give someone else a better idea.