Skip to content

Fix PPO entropy - #37

Merged
Ipsedo merged 47 commits into
masterfrom
develop
Sep 2, 2026
Merged

Fix PPO entropy#37
Ipsedo merged 47 commits into
masterfrom
develop

Conversation

@Ipsedo

@Ipsedo Ipsedo commented Sep 2, 2026

Copy link
Copy Markdown
Owner

Fix PPO entropy :

  • PID Lagrangian parameter for entropy bonus
  • constant target entropy
  • Some GUI update

Ipsedo added 30 commits August 26, 2026 00:21
* add const where it is semantically necessary => add tick method for enemy tank interface
* update tests regarding new behavior (first tick, then get_reward which is const now)
* add max shells != initial shells
…is 1.5s, hitting is +2s)

* shaping winner reward (nb_kills + 0.1 * nb_hits)
* shaping timeout penalty (1.0 - nb_steps / max_steps)
* Backtrack to constant timeout penalty and constant winner reward.
*To avoid policy collapse : start with discrete proba 50/50 and cosine annealing on target entropy from 0.98*max to 0.3*max
* miss_distance_scale from 1.5 to 3.0
* change hit reward from e^(-r^2) to 1/(1+r^2)
* start from lower continuous target entropy and sigma
* add aim reward and mu norm metrics
@Ipsedo
Ipsedo merged commit 9ade23f into master Sep 2, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant