{
"training_server_config": {
"config_update_polling_seconds": 10.0,
"default_hyperparameters": {
"DDPG": {
"seed": 1,
"gamma": 0.99,
"tau": 1e-2,
"learning_rate": 3e-3,
"batch_size": 128,
"buffer_size": 50000,
"learning_starts": 128,
"policy_frequency": 1,
"noise_scale": 0.1,
"train_iters": 50
},
"PPO": {
"discrete": true,
"seed": 0,
"traj_per_epoch": 1,
"clip_ratio": 0.1,
"gamma": 0.99,
"lam": 0.97,
"pi_lr": 3e-4,
"vf_lr": 3e-4,
"train_pi_iters": 40,
"train_v_iters": 40,
"target_kl": 0.01
},
"REINFORCE": {
"discrete": true,
"with_vf_baseline": true,
"seed": 1,
"traj_per_epoch": 8,
"gamma": 0.98,
"lam": 0.97,
"pi_lr": 3e-4,
"vf_lr": 1e-3,
"train_vf_iters": 80
},
"TD3": {
"seed": 1,
"gamma": 0.99,
"tau": 0.005,
"learning_rate": 3e-4,
"batch_size": 128,
"buffer_size": 50000,
"exploration_noise": 0.1,
"policy_noise": 0.2,
"noise_clip": 0.5,
"learning_starts": 25000,
"policy_frequency": 2
}
},
"training_tensorboard": {
"_comment1": "Runs `tensorboard --logdir /logs` in cwd on start up of server.",
"launch_tb_on_startup": true,
"_comment2": "scalar tags can be any column header from `progress.txt` files.",
"_comment3": "For more than one tag, separate by semi-colon (;)",
"scalar_tags": "AverageEpRet;LossQ",
"global_step_tag": "Epoch"
}
},
"transport_config": {
"addresses": {
"_comment": "gRPC uses only this address (prefix is unused).",
"model_server": {
"prefix": "tcp://",
"host": "127.0.0.1",
"port": "50051"
},
"trajectory_server": {
"prefix": "tcp://",
"host": "127.0.0.1",
"port": "7776"
},
"agent_listener": {
"prefix": "tcp://",
"host": "127.0.0.1",
"port": "7777"
},
"scaling_server": {
"prefix": "tcp://",
"host": "127.0.0.1",
"port": "7778"
},
"_comment2": "Only used when client-side inference is disabled or when client-side inference is used as fallback.",
"inference_server": {
"prefix": "tcp://",
"host": "127.0.0.1",
"port": "7779"
}
},
"local_model_module": {
"directory": "model_module",
"model_name": "training_model",
"format": "pt"
},
"max_traj_length": 100000000
}
}