},
'n_step': 1,
'evaluation_interval': None,
'evaluation_num_episodes': 1,
'timesteps_per_iteration': 1000,
'tau': 5e-3,
'buffer_size': 200000,
'prioritized_replay': False,
'optimization': {
'learning_rate': 5e-4,
'policy_loss_weight': 1.0,
'Q_loss_weight': 1.0,
'entropy_loss_weight': 1.0,
},
'learning_starts': 1000,
'sample_batch_size': 20,
'train_batch_size': 300,
'worker_side_prioritization': False,
'min_iter_time_s': 1,
}
main('SAC', configuration)
from paint_ppo import main
if __name__ == '__main__':
configuration = {
# 'model': {
# 'custom_model': 'paint_model',
# 'custom_options': {},
# },
'model': {
'fcnet_hiddens': [256, 128],
'use_lstm': False,
},
'num_workers': 15,
'sample_batch_size': 20,
}
main('A3C', configuration)
from paint_ppo import main
if __name__ == '__main__':
configuration = {
'num_workers': 15,
'num_gpus': 1,
'num_atoms': 1,
# 'v_min': -120.0,
# 'v_max': 120.0,
'dueling': True,
'double_q': True,
'hiddens': [256, 128],
'exploration_final_eps': 0.01,
"schedule_max_timesteps": 2000000,
'exploration_fraction': 0.2,
'timesteps_per_iteration': 1000,
'target_network_update_freq': 3000,
'soft_q': False,
'parameter_noise': False,
'batch_mode': 'truncate_episodes',
'buffer_size': 200000,
'prioritized_replay': True,
'compress_observations': False,
'learning_starts': 1000,
'sample_batch_size': 20,
'train_batch_size': 32,
}
main('APEX', configuration)
from paint_ppo import main
if __name__ == '__main__':
configuration = {
'num_workers': 15,
'twin_q': True,
'policy_delay': 2,
'smooth_target_policy': True,
# 'model': {
# # 'custom_model': 'paint_layer_model',
# # 'custom_options': {}, # extra options to pass to your model
# 'use_lstm': False,
# },
'actor_hiddens': [256, 128],
'critic_hiddens': [256, 128],
'timesteps_per_iteration': 1000,
'target_network_update_freq': 3000,
'tau': 1e-3,
'buffer_size': 200000,
'prioritized_replay': True,
'learning_starts': 1000,
'sample_batch_size': 20,
'train_batch_size': 32,
'num_gpus': 1,
# 'num_gpus_per_worker': 0,
}
main('APEX_DDPG', configuration)
},
'num_workers': 15,
'num_gpus': 1,
'batch_mode': 'truncate_episodes',
'observation_filter': 'NoFilter',
'lr': 0.0005,
'sample_batch_size': 50,
'train_batch_size': 750,
# 'num_sgd_iter': 16,
"num_data_loader_buffers": 4,
# how many train batches should be retained for minibatching. This conf
# only has an effect if `num_sgd_iter > 1`.
"minibatch_buffer_size": 4,
# set >0 to enable experience replay. Saved samples will be replayed with
# a p:1 proportion to new data samples.
# 'replay_proportion': 10,
# number of sample batches to store for replay. The number of transitions
# saved total will be (replay_buffer_num_slots * sample_batch_size).
# 'replay_buffer_num_slots': 100,
# level of queuing for sampling.
'max_sample_requests_in_flight_per_worker': 1,
'broadcast_interval': 3,
'grad_clip': 40.0,
'vf_loss_coeff': 0.5,
'entropy_coeff': 0.01,
}
main('IMPALA', configuration)