Hands-On Reinforcement Learning for Games

Noisy networks for exploration and importance sampling

Using noisy networks also introduces fuzziness in our action prediction. That is, since the weights of the network are now being pulled from a distribution that also means that they equally becoming distributional. We can also say they are stochastic and that stochasticity is defined by a distribution, basically meaning that the same input could yield two completely different results, which means we can no longer take just the max or best action because that is now just fuzzy. Instead, we need a way to decrease the size of the sampling distributions we use for weights and therefore the uncertainty we have in the actions the agent selects.

Decreasing the size of a distribution is more ...