Common Architecture

assigned to @dl337788

added Feature label

marked this issue as related to #11 (closed)

About half of the architecture outlined in "Playing Atari with Deep Reinforcement Learning" (Mnih et al., 2013) matches the architecture in https://github.com/openai/large-scale-curiosity/blob/master/utils.py#L133

More implementations including sources in: https://github.com/openai/baselines/blob/master/baselines/ppo1/cnn_policy.py#L22

added In Progress label

marked this issue as related to #7 (closed)

Logit distribution in https://github.com/openai/baselines/blob/master/baselines/common/distributions.py. Why no softmax?

The so called "Gumbel-max trick" is used to sample actions (in some implementations)

Might be the same as using the Categorical distribution from pytorch as shown in the Reinforcement Learning example

A larger CNN stack is described in "Human-level control through deep reinforcement learning" (Mnih et al., 2015)

added In Review label and removed In Progress label

marked this issue as related to #20 (closed)

mentioned in issue #20 (closed)

Tests in 3dbac4a4

closed

reopened

closed

removed In Review label

changed milestone to %Official dates

marked this issue as related to #54 (closed)