Reinforcement Learning: Actor-Critic Networks

In the previous blog post, we dived into the basic implementation of a deep Q-Learning Neural Network. It was a Policy-based duel-network which was used to learn the thief-police-gold game. Now, I have all of a sudden introduced two terms here: Policy-Based, Duel-Network. Policy-based methods are those which learn the probability distribution of the actions to take next of being in a given state. As it could be seen that we were using a softmax layer of output-size=number of possible actions, it was nothing but the policy-learning mechanism for the network to learn which action to take further depending on the probabilities. Moving on to the next term, Duel-Network. We had used two neural-networks in the previous blog, one of which was used and updated in an online manner while the other was updated (not so often) and used for predicting the policy values for the new_state. The reason for that was to be able to maintain some sort of consistency and not induce much randomness in the pol

In the previous blog post, we dived into the basic implementation of a deep Q-Learning Neural Network. It was a Policy-based duel-network which was used to learn the thief-police-gold game. Now, I have all of a sudden introduced two terms here: Policy-Based, Duel-Network. Policy-based methods are those which learn the probability distribution of the actions to take next of being in a given state. As it could be seen that we were using a softmax layer of output-size=number of possible actions, it was nothing but the policy-learning mechanism for the network to learn which action to take further