Playing Atari games with Deep Reinforcement Learning and Attention

DQN (Deep Q-learning) is a variant of Q-learning which is a method for solving Markov Decision Processes (MDP) by learning the value of actions from experiences. At every timestamp t, the environment gives the agent a state St which is an RGB frame, and reward Rt. Based on this, the agent chooses an action At. In return, the environment gives the agent a new reward Rt 1, and its new state St 1. And this cycle loops until the agent reaches a terminal state which marks the end of the episode. The agent's goal is to maximize the return Gt which is the sum of discounted rewards over an episode. The agent improves its performance by learning the value function Q(St, At) that predicts for every state-action pair (St, At) the future sum of discounted rewards using a deep neural network architecture to approximate Q. Reinforcement Learning agent DQN has been popularized by successful demonstrations on Atari games such as Pong[2]. To illustrate the problem, we will discuss state-of-the-art DQN using an Atari game as the environment.