mirror of
https://github.com/vale981/ray
synced 2025-03-09 12:56:46 -04:00
60d4d5e1aa
* Remove all __future__ imports from RLlib. * Remove (object) again from tf_run_builder.py::TFRunBuilder. * Fix 2xLINT warnings. * Fix broken appo_policy import (must be appo_tf_policy) * Remove future imports from all other ray files (not just RLlib). * Remove future imports from all other ray files (not just RLlib). * Remove future import blocks that contain `unicode_literals` as well. Revert appo_tf_policy.py to appo_policy.py (belongs to another PR). * Add two empty lines before Schedule class. * Put back __future__ imports into determine_tests_to_run.py. Fails otherwise on a py2/print related error.
53 lines
1.6 KiB
Python
53 lines
1.6 KiB
Python
"""A more stable successor to TD3.
|
|
|
|
By default, this uses a near-identical configuration to that reported in the
|
|
TD3 paper.
|
|
"""
|
|
|
|
from ray.rllib.agents.ddpg.ddpg import DDPGTrainer, \
|
|
DEFAULT_CONFIG as DDPG_CONFIG
|
|
from ray.rllib.utils import merge_dicts
|
|
|
|
TD3_DEFAULT_CONFIG = merge_dicts(
|
|
DDPG_CONFIG,
|
|
{
|
|
# largest changes: twin Q functions, delayed policy updates, and target
|
|
# smoothing
|
|
"twin_q": True,
|
|
"policy_delay": 2,
|
|
"smooth_target_policy": True,
|
|
"target_noise": 0.2,
|
|
"target_noise_clip": 0.5,
|
|
|
|
# other changes & things we want to keep fixed: IID Gaussian
|
|
# exploration noise, larger actor learning rate, no l2 regularisation,
|
|
# no Huber loss, etc.
|
|
"exploration_should_anneal": False,
|
|
"exploration_noise_type": "gaussian",
|
|
"exploration_gaussian_sigma": 0.1,
|
|
"learning_starts": 10000,
|
|
"pure_exploration_steps": 10000,
|
|
"actor_hiddens": [400, 300],
|
|
"critic_hiddens": [400, 300],
|
|
"n_step": 1,
|
|
"gamma": 0.99,
|
|
"actor_lr": 1e-3,
|
|
"critic_lr": 1e-3,
|
|
"l2_reg": 0.0,
|
|
"tau": 5e-3,
|
|
"train_batch_size": 100,
|
|
"use_huber": False,
|
|
"target_network_update_freq": 0,
|
|
"num_workers": 0,
|
|
"num_gpus_per_worker": 0,
|
|
"per_worker_exploration": False,
|
|
"worker_side_prioritization": False,
|
|
"buffer_size": 1000000,
|
|
"prioritized_replay": False,
|
|
"clip_rewards": False,
|
|
"use_state_preprocessor": False,
|
|
},
|
|
)
|
|
|
|
TD3Trainer = DDPGTrainer.with_updates(
|
|
name="TD3", default_config=TD3_DEFAULT_CONFIG)
|