Skip to content

Teach an agent how to play Atari Breakout with Deep Q-Learning and TensorFlow.

Notifications You must be signed in to change notification settings

wojciechmo/rl-dqn

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Deep Q-Learning

Teach RL agent how to play Atari Breakout with Deep Q-Learning and TensorFlow.

Deep Q-Learning uses neural network as Q-value function approximator. Search for Q-value function which satisfies Bellman equation: Q(s,a) = r + gamma * max[over a']Q(s',a'). Find optimal Q-value function recursively. During training force Bellman equation on sampled batches from replay buffer. No supervision available. Need to come up with labels G(s,a) for each (state, action, reward, next_state) training example.

Deep Q-Learning sketch:

Deep Q-Network architecture:

Deep Q-Learning is also stuffed with various stability tricks.

Use experience replay buffer to decorrelate training data by taking different parts from different trajectories instead of consecutive samples from single episode. Store tuples (s,a,r,s') using epsilon-greedy policy with respect to agent network. Update agent model using batches sampled from replay buffer.

Use target network to remove targets dependency on agent network parameters. Target network has the same architecture as agent network but different set of parameters. Copy agent network parameters to target network once in a while.

State is defined as 4 consecutive frames.

About

Teach an agent how to play Atari Breakout with Deep Q-Learning and TensorFlow.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages