Neural networks, including the popular "deep" networks, are useful tools for learning from data or from interactions. Typically, neural networks are trained with gradient descent. However, it is very hard to find gradients in very deep networks, or in the structure of the networks itself. Evolutionary algorithms can train neural network weights and structures through a process inspired by Darwinian evolution. I will give an introduction to the ideas behind neuroevolution, and discuss some applications in games, robotics and computer vision.