CODEWITHSUNDEEP
reinforcement-learning
T.L
T.L

Reputation: 21

Deep Reinforcement Learning with Atari Games: one DQN for all games or 49 DQNs for 49 games?

I would like to have a clarification about the article "Human level control through deep reinforcement learning" in Nature 2015. When I read it, I understand that they use a DQN with the same algo, network architecture and hyperparameters. Great! But they don't specify if they train each game from scratch and as a result we obtain one neural network per game (means 49 neural networks for the 49 games) or if they train all the game with a unique neural network (means only one neural network can play 49 games).

Is there someone who know what is the correct answer? Because it is not the same thing at all!:)

Thanks,

Upvotes: 0

Views: 381

Answers (1)

Afshin Oroojlooy
Afshin Oroojlooy

Reputation: 1434

In the paper that you mentioned, 49 networks are trained for 49 games: "A different network was trained on each game: the same network architecture, learning algorithm and hyperparameter settings (see Extended Data Table 1) were used across all games, showing that our approach is robust enough to work on a variety of games while incorporating only minimal prior knowledge", which is quoted from the paper.

There are algorithms which only train one network for all 49 games, e.g. https://arxiv.org/pdf/1809.04474.pdf

Upvotes: 1

Related Questions