Deep Reinforcement Learning has proved useful in solving tasks bounded by an environment/rules. Q-learning, a popular method of RL, has been shown to solve wide varieties of diverse environments. Due to the vigorous study/use of Q-learning, engineers and researchers have developed various add-ons and improvements to this algorithm. Using the once popular game 2048 as our environment, we will analyze these improvements to find the optimal Q-learning policy.
Playing Atari with Deep Reinforcement Learning, Google Deepmind:
https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf
Deep Reinforcement Learning with Double Q-learning, Google Deepmind:
https://arxiv.org/pdf/1509.06461.pdf
Rainbow: Combining Improvements in Deep Reinforcement Learning, Google Deepmind:
https://arxiv.org/pdf/1710.02298.pdf