def default_config() -> ClientConfig:
"""
:return: Default configuration for the experiment
"""
pg_agent_config = PolicyGradientAgentConfig(
gamma=1,
alpha_attacker=0.0001,
alpha_defender=0.0001,
epsilon=1,
render=False,
eval_sleep=0.9,
min_epsilon=0.01,
eval_episodes=100,
train_log_frequency=1,
epsilon_decay=0.9999,
video=True,
eval_log_frequency=1,
video_fps=5,
video_dir=default_output_dir() + "/results/videos",
num_episodes=100000000,
eval_render=False,
gifs=True,
gif_dir=default_output_dir() + "/results/gifs",
eval_frequency=250000,
attacker=True,
defender=True,
video_frequency=101,
save_dir=default_output_dir() + "/results/data",
checkpoint_freq=5000,
input_dim_attacker=(4 + 2) * 3,
output_dim_attacker=4 * 3,
input_dim_defender=(4 + 2) * 3,
output_dim_defender=5 * 3,
hidden_dim=32,
num_hidden_layers=2,
batch_size=2000,
gpu=True,
tensorboard=True,
tensorboard_dir=default_output_dir() + "/results/tensorboard",
optimizer="Adam",
lr_exp_decay=False,
lr_decay_rate=0.999,
state_length=1,
normalize_features=False,
merged_ad_features=True,
zero_mean_features=False,
gpu_id=0,
lstm_network=False,
lstm_seq_length=4,
num_lstm_layers=2,
optimization_iterations=10,
eps_clip=0.2,
max_gradient_norm=0.5,
gae_lambda=0.95)
env_name = "idsgame-v16"
client_config = ClientConfig(
env_name=env_name,
attacker_type=AgentType.PPO_OPENAI_AGENT.value,
defender_type=AgentType.PPO_OPENAI_AGENT.value,
mode=RunnerMode.TRAIN_DEFENDER_AND_ATTACKER.value,
pg_agent_config=pg_agent_config,
output_dir=default_output_dir(),
title="OpenAI-PPO vs OpenAI-PPO",
run_many=False,
random_seeds=[0, 999, 299, 399, 499])
# client_config = hp_tuning_config(client_config)
return client_config
def default_config() -> ClientConfig:
"""
:return: Default configuration for the experiment
"""
pg_agent_config = PolicyGradientAgentConfig(
gamma=0.999,
alpha_attacker=0.0001,
epsilon=1,
render=False,
eval_sleep=0.9,
min_epsilon=0.01,
eval_episodes=100,
train_log_frequency=1,
epsilon_decay=0.9999,
video=True,
eval_log_frequency=1,
video_fps=5,
video_dir=default_output_dir() + "/results/videos",
num_episodes=10000,
eval_render=False,
gifs=True,
gif_dir=default_output_dir() + "/results/gifs",
eval_frequency=1000000000,
attacker=True,
defender=False,
video_frequency=101,
save_dir=default_output_dir() + "/results/data",
checkpoint_freq=100,
input_dim_attacker=((4 + 2) * 4),
output_dim_attacker=(4 + 1) * 4,
input_dim_defender=((4 + 1) * 4),
output_dim_defender=5 * 4,
hidden_dim=128,
num_hidden_layers=2,
batch_size=64,
gpu=True,
tensorboard=True,
tensorboard_dir=default_output_dir() + "/results/tensorboard",
optimizer="Adam",
lr_exp_decay=False,
lr_decay_rate=0.999,
state_length=1,
normalize_features=False,
merged_ad_features=True,
zero_mean_features=False,
gpu_id=0,
lstm_network=False,
lstm_seq_length=4,
num_lstm_layers=2)
env_name = "idsgame-minimal_defense-v20"
client_config = ClientConfig(env_name=env_name,
attacker_type=AgentType.REINFORCE_AGENT.value,
mode=RunnerMode.TRAIN_ATTACKER.value,
pg_agent_config=pg_agent_config,
output_dir=default_output_dir(),
title="REINFORCE vs DefendMinimalDefender",
run_many=False,
random_seeds=[0, 999, 299, 399, 499],
random_seed=399)
#client_config = hp_tuning_config(client_config)
return client_config
def default_config() -> ClientConfig:
"""
:return: Default configuration for the experiment
"""
pg_agent_config = PolicyGradientAgentConfig(
gamma=1,
alpha_attacker=0.0001,
epsilon=1,
render=False,
alpha_defender=0.0001,
eval_sleep=0.9,
min_epsilon=0.01,
eval_episodes=1000,
train_log_frequency=1,
epsilon_decay=0.9999,
video=True,
eval_log_frequency=500,
video_fps=5,
video_dir=default_output_dir() + "/results/videos",
num_episodes=100000000,
eval_render=False,
gifs=True,
gif_dir=default_output_dir() + "/results/gifs",
eval_frequency=55000,
attacker=False,
defender=True,
video_frequency=1001,
save_dir=default_output_dir() + "/results/data",
checkpoint_freq=50,
input_dim_attacker=((4 + 2) * 4),
output_dim_attacker=(4 + 1) * 4,
input_dim_defender=((4 + 1) * 4),
output_dim_defender=5 * 4,
hidden_dim=128,
num_hidden_layers=3,
pi_hidden_layers=1,
pi_hidden_dim=128,
vf_hidden_layers=1,
vf_hidden_dim=128,
batch_size=2000,
gpu=False,
tensorboard=True,
tensorboard_dir=default_output_dir() + "/results/tensorboard",
optimizer="Adam",
lr_exp_decay=False,
lr_decay_rate=0.999,
state_length=1,
normalize_features=False,
merged_ad_features=True,
zero_mean_features=False,
gpu_id=0,
lstm_network=False,
lstm_seq_length=4,
num_lstm_layers=2,
optimization_iterations=10,
eps_clip=0.2,
max_gradient_norm=0.5,
gae_lambda=0.95,
cnn_feature_extractor=False,
features_dim=512,
flatten_feature_planes=False,
cnn_type=5,
vf_coef=0.5,
ent_coef=0.001,
render_attacker_view=False,
lr_progress_power_decay=4,
lr_progress_decay=True,
use_sde=False,
sde_sample_freq=4,
one_hot_obs=False,
lstm_core=False,
lstm_hidden_dim=32,
multi_channel_obs=False,
channel_1_dim=32,
channel_1_layers=2,
channel_1_input_dim=16,
channel_2_dim=32,
channel_2_layers=2,
channel_2_input_dim=16,
channel_3_dim=32,
channel_3_layers=2,
channel_3_input_dim=4,
channel_4_dim=32,
channel_4_layers=2,
channel_4_input_dim=4,
mini_batch_size=64,
ar_policy=True,
attacker_node_input_dim=((4 + 2) * 4),
attacker_at_net_input_dim=(4 + 2),
attacker_at_net_output_dim=(4 + 1),
attacker_node_net_output_dim=4,
defender_node_input_dim=((4 + 1) * 4),
defender_at_net_input_dim=(4 + 1),
defender_node_net_output_dim=4,
defender_at_net_output_dim=5)
env_name = "idsgame-maximal_attack-v19"
client_config = ClientConfig(
env_name=env_name,
defender_type=AgentType.PPO_OPENAI_AGENT.value,
mode=RunnerMode.TRAIN_DEFENDER.value,
pg_agent_config=pg_agent_config,
output_dir=default_output_dir(),
title="AttackMaximalAttacker vs OpenAI-PPO",
run_many=False,
random_seeds=[0, 999, 299, 399, 499])
# client_config = hp_tuning_config(client_config)
return client_config
def default_config() -> ClientConfig:
"""
:return: Default configuration for the experiment
"""
opponent_pool_config = OpponentPoolConfig(pool_maxsize=100000,
pool_increment_period=50,
head_to_head_period=1,
quality_scores=True,
quality_score_eta=0.01,
initial_quality=1000,
pool_prob=0.5)
pg_agent_config = PolicyGradientAgentConfig(
gamma=0.999,
alpha_attacker=0.0001,
alpha_defender=0.0001,
epsilon=1,
render=False,
eval_sleep=0.9,
min_epsilon=0.01,
eval_episodes=100,
train_log_frequency=100,
epsilon_decay=0.9999,
video=True,
eval_log_frequency=1,
video_fps=5,
video_dir=default_output_dir() + "/results/videos",
num_episodes=450001,
eval_render=False,
gifs=True,
gif_dir=default_output_dir() + "/results/gifs",
eval_frequency=20000,
attacker=True,
defender=True,
video_frequency=101,
save_dir=default_output_dir() + "/results/data",
checkpoint_freq=5000,
input_dim_attacker=((4 + 2) * 4),
output_dim_attacker=(4 + 1) * 4,
input_dim_defender=((4 + 1) * 4),
output_dim_defender=5 * 4,
hidden_dim=128,
num_hidden_layers=2,
batch_size=64,
gpu=False,
tensorboard=True,
tensorboard_dir=default_output_dir() + "/results/tensorboard",
optimizer="Adam",
lr_exp_decay=False,
lr_decay_rate=0.99999,
normalize_features=False,
merged_ad_features=True,
zero_mean_features=False,
gpu_id=0,
opponent_pool_config=opponent_pool_config,
alternating_optimization=50,
opponent_pool=True,
baselines_in_pool=True,
alternating_period=50)
env_name = "idsgame-v20"
client_config = ClientConfig(
env_name=env_name,
attacker_type=AgentType.REINFORCE_AGENT.value,
defender_type=AgentType.REINFORCE_AGENT.value,
mode=RunnerMode.TRAIN_DEFENDER_AND_ATTACKER.value,
pg_agent_config=pg_agent_config,
output_dir=default_output_dir(),
title="REINFORCE vs REINFORCE",
run_many=False,
random_seeds=[0, 999, 299, 399, 499],
random_seed=0)
#client_config = hp_tuning_config(client_config)
return client_config
def default_config() -> ClientConfig:
"""
:return: Default configuration for the experiment
"""
env_name = "idsgame-v18"
pg_agent_config = PolicyGradientAgentConfig(
gamma=1,
alpha_attacker=0.00001,
epsilon=1,
render=False,
alpha_defender=0.0001,
eval_sleep=0.9,
min_epsilon=0.01,
eval_episodes=100,
train_log_frequency=1,
epsilon_decay=0.9999,
video=True,
eval_log_frequency=1,
video_fps=5,
video_dir=default_output_dir() + "/results/videos",
num_episodes=100000000,
eval_render=False,
gifs=True,
gif_dir=default_output_dir() + "/results/gifs",
eval_frequency=100000,
attacker=True,
defender=False,
video_frequency=101,
save_dir=default_output_dir() + "/results/data",
checkpoint_freq=5000,
input_dim_attacker=(4 + 2) * 2,
output_dim_attacker=(4 + 1) * 2,
input_dim_defender=(4 + 2) * 3,
output_dim_defender=5 * 3,
hidden_dim=64,
num_hidden_layers=4,
batch_size=2000,
gpu=False,
tensorboard=True,
tensorboard_dir=default_output_dir() + "/results/tensorboard",
optimizer="Adam",
lr_exp_decay=False,
lr_decay_rate=0.999,
state_length=1,
normalize_features=False,
merged_ad_features=True,
zero_mean_features=False,
gpu_id=0,
lstm_network=False,
lstm_seq_length=4,
num_lstm_layers=2,
optimization_iterations=10,
eps_clip=0.2,
max_gradient_norm=0.5,
gae_lambda=0.95,
cnn_feature_extractor=False,
features_dim=512,
flatten_feature_planes=False,
attacker_load_path=
"/home/kim/storage/workspace/gym-idsgame/experiments/manual_play/v18/minimal_defense/manual_vs_openai_ppo/1591093705.5003314_attacker_policy_network.zip"
)
client_config = ClientConfig(
env_name=env_name,
attacker_type=AgentType.PPO_OPENAI_AGENT.value,
mode=RunnerMode.MANUAL_DEFENDER.value,
output_dir=default_output_dir(),
title="OpenAI PPO vs ManualDefender",
pg_agent_config=pg_agent_config,
bot_attacker=True)
return client_config
def default_config() -> ClientConfig:
"""
:return: Default configuration for the experiment
"""
env_name = "idsgame-v19"
pg_agent_config = PolicyGradientAgentConfig(
gamma=1,
alpha_attacker=0.0001,
epsilon=1,
render=False,
alpha_defender=0.0001,
eval_sleep=0.9,
min_epsilon=0.01,
eval_episodes=1000,
train_log_frequency=1,
epsilon_decay=0.9999,
video=True,
eval_log_frequency=500,
video_fps=5,
video_dir=default_output_dir() + "/results/videos",
num_episodes=100000000,
eval_render=False,
gifs=True,
gif_dir=default_output_dir() + "/results/gifs",
eval_frequency=55000,
attacker=False,
defender=True,
video_frequency=1001,
save_dir=default_output_dir() + "/results/data",
checkpoint_freq=250,
input_dim_attacker=((4 + 2) * 4),
output_dim_attacker=(4 + 1) * 4,
input_dim_defender=((4 + 1) * 4),
output_dim_defender=5 * 4,
hidden_dim=128,
num_hidden_layers=2,
pi_hidden_layers=1,
pi_hidden_dim=128,
vf_hidden_layers=1,
vf_hidden_dim=128,
batch_size=2000,
gpu=False,
tensorboard=True,
tensorboard_dir=default_output_dir() + "/results/tensorboard",
optimizer="Adam",
lr_exp_decay=False,
lr_decay_rate=0.999,
state_length=1,
normalize_features=False,
merged_ad_features=True,
zero_mean_features=False,
gpu_id=0,
lstm_network=False,
lstm_seq_length=4,
num_lstm_layers=2,
optimization_iterations=10,
eps_clip=0.2,
max_gradient_norm=0.5,
gae_lambda=0.95,
cnn_feature_extractor=False,
features_dim=512,
flatten_feature_planes=False,
cnn_type=5,
vf_coef=0.5,
ent_coef=0.001,
render_attacker_view=False,
lr_progress_power_decay=4,
lr_progress_decay=True,
use_sde=False,
sde_sample_freq=4,
one_hot_obs=False,
lstm_core=False,
lstm_hidden_dim=32,
multi_channel_obs=False,
channel_1_dim=32,
channel_1_layers=2,
channel_1_input_dim=16,
channel_2_dim=32,
channel_2_layers=2,
channel_2_input_dim=16,
channel_3_dim=32,
channel_3_layers=2,
channel_3_input_dim=4,
channel_4_dim=32,
channel_4_layers=2,
channel_4_input_dim=4,
mini_batch_size=64,
ar_policy=True,
attacker_node_input_dim=((4 + 2) * 4),
attacker_at_net_input_dim=(4 + 2),
attacker_at_net_output_dim=(4 + 1),
attacker_node_net_output_dim=4,
defender_node_input_dim=((4 + 1) * 4),
defender_at_net_input_dim=(4 + 1),
defender_node_net_output_dim=4,
defender_at_net_output_dim=5,
defender_load_path=
"/home/kim/workspace/gym-idsgame/experiments/manual_play/v19/maximal_attack/manual_vs_openai_ppo/1592125075.4390159_defender_node_policy_network.zip"
)
client_config = ClientConfig(
env_name=env_name,
defender_type=AgentType.PPO_OPENAI_AGENT.value,
mode=RunnerMode.MANUAL_ATTACKER.value,
output_dir=default_output_dir(),
title="ManualAttacker vs OpenAI PPO",
pg_agent_config=pg_agent_config,
bot_defender=True)
return client_config
pg_agent_config = PolicyGradientAgentConfig(
gamma=1,
alpha_attacker=0.0001,
epsilon=1,
render=False,
alpha_defender=0.0001,
eval_sleep=0.9,
min_epsilon=0.01,
eval_episodes=1000,
train_log_frequency=args.train_log_frequency,
epsilon_decay=0.9999,
video=False,
eval_log_frequency=args.eval_log_frequency,
video_fps=5,
video_dir=default_output_dir() + "/results/videos",
num_episodes=args.num_episodes,
eval_render=False,
gifs=False,
gif_dir=default_output_dir() + "/results/gifs/" + args.experiment_id,
eval_frequency=args.eval_frequency,
attacker=args.attacker,
defender=args.defender,
video_frequency=1001,
save_dir=default_output_dir() + "/results/data/" + args.experiment_id,
checkpoint_freq=250,
input_dim_attacker=((4 + 2) * 4),
output_dim_attacker=(4 + 1) * 4,
input_dim_defender=((4 + 1) * 4),
output_dim_defender=5 * 4,
hidden_dim=128,
num_hidden_layers=2,
pi_hidden_layers=1,
pi_hidden_dim=128,
vf_hidden_layers=1,
vf_hidden_dim=128,
batch_size=2000,
gpu=False,
tensorboard=True,
tensorboard_dir=default_output_dir() + "/results/tensorboard/" +
args.experiment_id,
optimizer="Adam",
lr_exp_decay=False,
lr_decay_rate=0.999,
state_length=1,
normalize_features=False,
merged_ad_features=True,
zero_mean_features=False,
gpu_id=0,
lstm_network=False,
lstm_seq_length=4,
num_lstm_layers=2,
optimization_iterations=10,
eps_clip=0.2,
max_gradient_norm=0.5,
gae_lambda=0.95,
cnn_feature_extractor=False,
features_dim=512,
flatten_feature_planes=False,
cnn_type=5,
vf_coef=0.5,
ent_coef=0.001,
render_attacker_view=True,
lr_progress_power_decay=4,
lr_progress_decay=True,
use_sde=False,
sde_sample_freq=4,
one_hot_obs=False,
lstm_core=False,
lstm_hidden_dim=32,
multi_channel_obs=False,
channel_1_dim=32,
channel_1_layers=2,
channel_1_input_dim=16,
channel_2_dim=32,
channel_2_layers=2,
channel_2_input_dim=16,
channel_3_dim=32,
channel_3_layers=2,
channel_3_input_dim=4,
channel_4_dim=32,
channel_4_layers=2,
channel_4_input_dim=4,
mini_batch_size=64,
ar_policy=True,
attacker_node_input_dim=((4 + 2) * 4),
attacker_at_net_input_dim=(4 + 2),
attacker_at_net_output_dim=(4 + 1),
attacker_node_net_output_dim=4)
