What i mean is games like chess, draughts, tic tac toe, 2048, Super Mario?, in general games that require multiple plays, moves to complete.
I'm pretty sure one could use Genetic Algorithms, but i'm willing to know if there's a way to train it with pure backpropagation. What would I use as target values?
I was thinking of letting him play a number of moves until the game ends and use the end result as target values, but i think that would only train the last move, won't it?
The other approach would be to train every move, but then what would the target values be? You can have multiple valid choices move-by-move, but surely enough one is better on the long-term. How to choose it without me trial-and-error'ing it out?
Is it just not possible to achieve with Backprop?