}
maps = []
# Symmetric (6)
maps += ["3m"] #, "8m", "25m", "2s3z", "3s5z", "MMM"]
# Asymmetric (6)
# maps += ["5m_6m", "8m_9m", "10m_11m", "27m_30m"]
# maps += ["MMM2", "3s5z_3s6z"]
# Micro (10)
# maps += ["3s_vs_3z", "3s_vs_4z", "3s_vs_5z"]
# maps += ["micro_2M_Z"]
# maps += ["micro_baneling"]
# maps += ["micro_colossus"]
# maps += ["micro_corridor"]
# maps += ["micro_focus"]
# maps += ["micro_retarget"]
# maps += ["micro_bane"]
for map_name in maps:
name = label
extend_param_dicts(param_dicts,
shared_params, {
"name": name,
"env_args.map_name": map_name
},
repeats=parallel_repeat)
"env_args.steps": 10,
"env_args.good_branches": 2,
"batch_size_run": 1,
"test_interval": 1000,
"test_nepisode": 64,
"test_greedy": True,
"log_interval": 1000,
"runner_log_interval": 2000,
"learner_log_interval": 2000,
"buffer_cpu_only": True, # 5k buffer is too big for VRAM!
"buffer_size": 1000,
"epsilon_finish": 0.01,
"epsilon_anneal_time": 500,
"discrim_size": 32,
}
name = "noisemix"
extend_param_dicts(param_dicts, shared_params,
{
"name": name,
"noise_dim": [16],
"bandit_iters": 100,
"noise_bandit": [True],
"rnn_discrim": [True],
"mi_loss": [1],
# "entropy_scaling": [0.001, 0.01, 0.1]
},
repeats=parallel_repeat)
# "batch_size_run": 1,
"test_interval": 30000,
"test_nepisode": 8,
"test_greedy": True,
"log_interval": 30000,
"runner_log_interval": 30000,
"learner_log_interval": 30000,
"buffer_cpu_only": True, # 5k buffer is too big for VRAM!
"buffer_size": 3000,
"epsilon_finish": 0.05,
"epsilon_anneal_time": 250000,
#"discrim_size": 32,
}
name = "noisemix"
extend_param_dicts(
param_dicts,
shared_params,
{
"env_args.map_name": ["2_corridors"],
"name": name,
"noise_dim": [16],
#"bandit_iters": 100,
"noise_bandit": [True],
"rnn_discrim": [True],
"mi_loss": [0.001],
"entropy_scaling": [0.001]
},
repeats=parallel_repeat)
"test_interval": 30000,
"test_nepisode": 8,
"test_greedy": True,
"save_model": False,
#"save_model_interval": 250 * 1000,
"log_interval": 30000,
"runner_log_interval": 30000,
"learner_log_interval": 30000,
"buffer_cpu_only": True, # 5k buffer is too big for VRAM!
# "training_iters": 1,
"buffer_size": 3000
}
name = label + config + "_" + env_config
extend_param_dicts(param_dicts,
shared_params, {
"lr": [0.0005],
"epsilon_anneal_time": [250000],
"env_args.reward_only_positive": [False],
"env_args.reward_negative_scale": [1.0],
"env_args.map_name": ["8m", "MMM2", "10m_vs_11m"],
"name": name,
"rnn_hidden_dim": [128],
"obs_agent_id": [False],
"grad_norm_clip": [10],
"target_update_mode": ["soft"],
"target_update_tau": [0.05],
"mi_loss": [0.001, 0.01],
"agent": ["noise_rnn_deep"]
},
repeats=parallel_repeat)