This blog is a memo for reinforcement learning, inspired by an introductory note. I would suggest people read the original post. This memo will give an outline for its thoughts and logic. This post lacks technical details. Referring to the paper listed at the end may be useful.


As the work by DeepMind stated, they’ve use reinforcement learning to play a game. If we want to teach machine to play a game, we may take it as a classification task. Given pictures of a screen, the model output an action. But this is not how we humans learn.

When we’re playing a game, we choose a lot of actions to get a final reward. How a reward is related to each previous actions is the problem we are facing. That’s known as credit assignment problem.

Another question is how much will an agent strive for the reward. Will it just try once and give up with a reward of very low points? This is explore-exploit dilemma.


Often we use Markov Process to model a game. The agent in specific state of the game environment must choose actions to get the reward, in special policy. The states in all therefore forms a Markov Process.

Total reward will be .

At time , the total reward from now on will be .

Because the more we move towards into the future, the more likely the reward will be converge. The total reward will be discounted.

A good policy would be to maximize the future reward.


Let Q function be , which is the future reward we perform action a in state s and continue optimally from then on.

At time t, also define the chosen action as .

As stated above, we compute Q function iteratively using the maximum future reward.

parameter alpha is something like the learning rate.

Deep Q-Learning

But if the number of states is too much, we may turn to DNN for a better generalization.

Like the structure above, we may use the right model since it computes the future reward for all action at once.

Experience Replay

Experience replay is a trick. We store all experience in the special memory, and use random minibatches from it to training. Maybe less likely will we go into local minimum.


Use another -greedy exploration technique. Use a probability to move randomly, which will decrease from 1 to 0.


  1. Blog post: Demystifying Deep Reinforcement Learning by nervana.
  2. Reinforcement Learning course by David Silver.
  3. Reinforcement Learning course from UC Berkeley.
  4. Q-Learning detailed blog
  5. Mnih, Volodymyr, et al. “Playing atari with deep reinforcement learning.” arXiv preprint arXiv:1312.5602 (2013).