Suppose an AI is to play the game flappy bird. And the fitness function is how long the bird has traveled before the game ends.
Would we have multiple neural networks initialized at the beginning with random weights (as in each bird has its own network) and then we determine the neural networks that have lasted the longest for the game and then we perform a selection of weights from the "better" neural networks followed by mutation? Those will then be used as the new weights of a brand new neural network (ie the offspring from two "better" neural networks?)?
If that is the case, does that mean there is no backpropagation because there isn't a cost function?