The very first question seemed to iffy with me. Was said that if a AI played only 10 games over 1000 it would get worse, not better. I would still say it would get better, just not as much…right. Doing games/task over and over will cause a greater gain that is indeed true. However, doing a smaller amount would still lead to learning and thus still is a gain right? The idea comes from my Philosophy thought. Unless the AI was forced to make incorrect moves, it will always learn. It’s just the more times the more it will learn. So I think a 3rd option should be there as a “It would get better, just not as good”
Hello @Samad_Khan! Welcome to our community!
It will learn from 10 games but we cannot say it would be better. Playing 10 games multiple times is not making it better. To become better, need to try different games.
What is the background of your question?
Best,
Saif.
However, It will definitely won’t get worse. It will remain on a generally same level, or get a tiny bit better, even if it is on an EXTREMELY small amount. That’s was my point. Sure 10 games are not enough to learn much from, was with things are now, it is indeed able to learn a tiny bit. See how “gets worse” can seem like a more wrong answer?
Getting better or worse is depend on the type of game or task it is trained on. If a five-year kid can play better than the ML model (trained on 10 games only), then what do you think is the performance of a model? Better or worse?
Saif.
A kid with no experience playing simply one game will gain knowledge on it, not lose knowledge on it. That is the point. If a machine is at 0 and does simply 1 game, it may be at 0.00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001 (probably more zeros) now. It stilled is going to learn from the game.
Yes, it will be learning.
The solution surfaces for Neural Network cost functions are extremely complex. There is no guarantee that every step of Gradient Descent will always lead to a better solution. If you run a bunch of experiments of this type and plot the cost function, you will see that the progress is not guaranteed to be monotonic. So over any small number of iterations there are no guarantees you can make. It’s also possible that you’ve made errors in how you have chosen various hyperparameters and maybe you will never get to a good solution without some further tuning. Not every experiment you run will be successful. 
But as Saif said earlier, it would be helpful to know which lecture in which course was the starting point for your question. You’ve filed this under General Discussion and not given us any context really.