Deep Q-Learning since Q-learning involves learning estimates from estimates, such overestimation can be problematic.
The solution involves using two separate Q-value estimators, each of which is used to update the other. Using these independent estimators, we can unbiased Q-value estimates of the actions selected using the opposite estimator. We can thus avoid maximization bias by disentangling our updates from biased estimates.
Therefore Double Deep Q-Learning is been invented to fix this issue.
- https://medium.com/@qempsil0914/deep-q-learning-part2-double-deep-q-network-double-dqn-b8fc9212bbb2
- https://ai.stackexchange.com/questions/21515/is-there-any-good-reference-for-double-deep-q-learning
- https://unnatsingh.medium.com/deep-q-network-with-pytorch-d1ca6f40bfda
- https://aamrani1999.medium.com/double-deep-q-networks-f6cf5b1b249
- https://towardsdatascience.com/double-deep-q-networks-905dd8325412
- https://github.com/chinancheng/DDQN.pytorch
- https://github.com/germain-hug/Deep-RL-Keras