The great bug hunt of 2020

assigned to @dl337788

changed milestone to %Official dates

added Bug label

Final performance on Breakout with 1 run is 104.97; reward graph shows peaks of up to 400 episode reward. However, the algorithm is very unstable. Final performance in PPO paper is 274.8, which should be achievable if learning becomes more stable.

TODO:

• check loss calculation using open source implementations available on GitHub
• check hyperparameters, most importantly batch size, rollout length and number of updates

marked this issue as related to #71 (closed)

marked this issue as related to #73 (closed)

added In Progress label

marked this issue as related to #78 (closed)

marked this issue as related to #79 (closed)

marked this issue as related to #80 (closed)

See:

• https://www.reddit.com/r/reinforcementlearning/comments/d3wym2/catastrophic_unlearning_in_ppo_a_plausible_cause/
• https://www.reddit.com/r/reinforcementlearning/comments/bqh01v/having_trouble_with_ppo_rewards_crashing/

marked this issue as related to #49 (closed)

marked this issue as related to #82 (closed)

See also: https://github.com/llSourcell/Unity_ML_Agents/blob/master/docs/best-practices-ppo.md

marked this issue as related to #98 (closed)

Closed with a1fa87e6

closed

removed In Progress label