Hands-On Intelligent Agents with OpenAI Gym

Using neural networks to approximate Q-functions

Neural networks are shown to be effective as universal function approximators. In fact, there is a universal approximation theorem that states that a single hidden layer feed-forward neural network can approximate any continuous function that is closed and bounded in . It basically means that even simple (shallow) neural networks can approximate several functions. Doesn't it feel too good to be true that you can use a simple neural network with a fixed number of weights/parameters to approximate practically any function? It is actually true, except for one note that prevents it from being used ...