Closed gal-leibovich closed 5 years ago
The TD3 implementation in this repo significantly outperforms the paper published results (although Coach uses v2 environments and the official TD3 repo uses v1 environments, the performance between the two should be similar, e.g. as stated and shown here.
Env | TD3 (Coach) | TD3 (official repo) |
---|---|---|
Hopper | 3628+- 80 | 3564 +- 115 |
Walker2D | 4766 +- 800 | 4682 +- 540 |
Reacher | -3.468 +- 0.35 | -3.6 +- 0.56 |
Half Cheetah | 10888 +- 1250 | 9637 +- 859 |
Ant | 5090 +- 1000 | 4372 +- 1000 |
Comparison of TD3 (called FixedCriticInputBugAndBootstrapBug) to DDPG (no batchnorm)
Hopper
Walker2D
Reacher
Half Cheetah
Ant
Ready for review - reproducing and even surpassing paper results.
Note: Slight renaming of the
noise_percentage_schedule
parameter resulted in some changes in unrelated files.Also, some small fixes to DDPG.