Asynchronous Methods for Deep Reinforcement Learning (A3C)

After training for 6 hours

• A3C implementation in TensorFlow
• There are two versions
  • A3C using threads
  • A3C using distributed tensorflow
• Tested on "Pong-v0" which is a stochastic environment due to random frame skipping
• Original paper: https://arxiv.org/abs/1602.01783

Tensorboard

Episode length

• The time length of an episode increases as it starts to play neck and neck against the AI
• After an hour, the trained agent begins to beat the AI easily

Episode reward

• The maximum score possible is 21