Repo for this ArXiv paper. All training scripts are located in the scripts/ directory. To be updated with more extensive experiments on different agents (e.g. SAC, PPO).
Hyperparameters to consider (as of Dec 2022)
| Hyperparameter | values |
|---|---|
learning_rate |
1e-3, 5e-4 |
rho (perturbation margin) |
3, 5, 7, 10, 20% |
rho_sample |
10, 20, 30 |
lambda (look ahead steps) |
1, 10 |
sample_heuristics |
max, mode of top percentile |
model-free algo |
DQN (Actor Critic, SAC, PPO) |