Hands-On Intelligent Agents with OpenAI Gym

Noisy nets

If you recall, we used an -greedy policy for the deep Q-learning agent in Chapter 6, Implementing an Intelligent Agent for Optimal Discrete Control Using Deep Q-Learning, to take action based on the action-values learned by the deep Q-network, which basically means taking the action with the highest action-value for a given state most of the time, except when, for some tiny fraction of the time (that is, with a very small probability ), the agent selects a random action. This may prevent the agent from exploring more reward states, ...