Bart/bullet3
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Should sort best directions that gives the highest rewards

Browse Source 
Found bug where the sort order would incorporate the lowest N rewards, instead of the highest. With this fix, one can now use fewer best directions instead.
This commit is contained in:
YungKC
2019-04-16 12:20:08 -07:00
committed by GitHub
parent bf0eb4703d
commit 61987dea85
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
examples/pybullet/gym/pybullet_envs/ARS/ars.py
View File

@@ -26,7 +26,7 @@ class Hp():
self.episode_length = 1000 self.episode_length = 1000
self.learning_rate = 0.02 self.learning_rate = 0.02
self.nb_directions = 16 self.nb_directions = 16
self.nb_best_directions = 16 self.nb_best_directions = 8
assert self.nb_best_directions <= self.nb_directions assert self.nb_best_directions <= self.nb_directions
self.noise = 0.03 self.noise = 0.03
self.seed = 1 self.seed = 1
@@ -194,7 +194,7 @@ def train(env, policy, normalizer, hp, parentPipes, args):
# Sorting the rollouts by the max(r_pos, r_neg) and selecting the best directions # Sorting the rollouts by the max(r_pos, r_neg) and selecting the best directions
scores = {k:max(r_pos, r_neg) for k,(r_pos,r_neg) in enumerate(zip(positive_rewards, negative_rewards))} scores = {k:max(r_pos, r_neg) for k,(r_pos,r_neg) in enumerate(zip(positive_rewards, negative_rewards))}
order = sorted(scores.keys(), key = lambda x:scores[x])[:hp.nb_best_directions] order = sorted(scores.keys(), key = lambda x:-scores[x])[:hp.nb_best_directions]
rollouts = [(positive_rewards[k], negative_rewards[k], deltas[k]) for k in order] rollouts = [(positive_rewards[k], negative_rewards[k], deltas[k]) for k in order]
# Updating our policy # Updating our policy