from ray.rllib.utils import merge_dicts
APEX_DDPG_DEFAULT_CONFIG = merge_dicts(
DDPG_CONFIG, # see also the options in ddpg.py, which are also supported
{
"optimizer": merge_dicts(
DDPG_CONFIG["optimizer"], {
"max_weight_sync_delay": 400,
"num_replay_buffer_shards": 4,
"debug": False
}),
"n_step": 3,
"num_gpus": 0,
"num_workers": 32,
"buffer_size": 2000000,
"learning_starts": 50000,
"train_batch_size": 512,
"sample_batch_size": 50,
"target_network_update_freq": 500000,
"timesteps_per_iteration": 25000,
"per_worker_exploration": True,
"worker_side_prioritization": True,
"min_iter_time_s": 30,
},
)
ApexDDPGTrainer = DDPGTrainer.with_updates(
name="APEX_DDPG",
default_config=APEX_DDPG_DEFAULT_CONFIG,
**APEX_TRAINER_PROPERTIES)
# 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)
},
# other changes & things we want to keep fixed:
# larger actor learning rate, no l2 regularisation, no Huber loss, etc.
"learning_starts": 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,
"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,
get_policy_class=None,
)
APEX_DDPG_DEFAULT_CONFIG = DDPGTrainer.merge_trainer_configs(
DDPG_CONFIG, # see also the options in ddpg.py, which are also supported
{
"optimizer": {
"max_weight_sync_delay": 400,
"num_replay_buffer_shards": 4,
"debug": False
},
"exploration_config": {
"type": "PerWorkerOrnsteinUhlenbeckNoise"
},
"n_step": 3,
"num_gpus": 0,
"num_workers": 32,
"buffer_size": 2000000,
"learning_starts": 50000,
"train_batch_size": 512,
"rollout_fragment_length": 50,
"target_network_update_freq": 500000,
"timesteps_per_iteration": 25000,
"worker_side_prioritization": True,
"min_iter_time_s": 30,
},
)
ApexDDPGTrainer = DDPGTrainer.with_updates(
name="APEX_DDPG",
default_config=APEX_DDPG_DEFAULT_CONFIG,
execution_plan=apex_execution_plan)
*,
explore=True,
is_training=False,
**kwargs):
model_out, _ = model({
"obs": obs_batch,
"is_training": is_training,
}, [], None)
dist_inputs = model.get_policy_output(model_out)
dist_class, logit_dim = ModelCatalog.get_action_dist(
model.action_space, policy.config["model"], framework="torch")
return dist_inputs, dist_class, [] # []=state out
PADDPGTorchPolicy = DDPGTorchPolicy.with_updates(
# loss_fn=paddpg_loss,
# action_sampler_fn=action_sampler_fn,
validate_spaces=None,
action_distribution_fn=get_distribution_inputs_and_class,
make_model_and_action_dist=None,
make_model=build_paddpg_models,
)
def get_policy_class(config):
return PADDPGTorchPolicy
PADDPGTrainer = DDPGTrainer.with_updates(
name="PADDPG",
default_config=DDPG_CONFIG,
get_policy_class=get_policy_class,
default_policy=PADDPGTorchPolicy,
)