def train_attacker(
config: ClientConfig) -> Union[ExperimentResult, ExperimentResult]:
"""
Trains an attacker agent in the environment
:param config: Training configuration
:return: trainresult, evalresult
"""
env: IdsGameEnv = None
env = gym.make(config.env_name,
idsgame_config=config.idsgame_config,
save_dir=config.output_dir + "/results/data/" +
str(config.random_seed),
initial_state_path=config.initial_state_path)
if config.title is not None:
env.idsgame_config.render_config.title = config.title
attacker: TrainAgent = None
if config.attacker_type == AgentType.TABULAR_Q_AGENT.value:
attacker = TabularQAgent(env, config.q_agent_config)
elif config.attacker_type == AgentType.DQN_AGENT.value:
attacker = DQNAgent(env, config.q_agent_config)
elif config.attacker_type == AgentType.REINFORCE_AGENT.value:
attacker = ReinforceAgent(env, config.pg_agent_config)
elif config.attacker_type == AgentType.ACTOR_CRITIC_AGENT.value:
attacker = ActorCriticAgent(env, config.pg_agent_config)
elif config.attacker_type == AgentType.BAYES_ACTOR_CRITIC_AGENT.value:
attacker = BayesActorCriticAgent(env, config.pg_agent_config)
elif config.attacker_type == AgentType.PPO_AGENT.value:
attacker = PPOAgent(env, config.pg_agent_config)
elif config.attacker_type == AgentType.PPO_OPENAI_AGENT.value:
wrapper_env = BaselineEnvWrapper(
config.env_name,
idsgame_config=config.idsgame_config,
save_dir=config.output_dir + "/results/data/" +
str(config.random_seed),
initial_state_path=config.initial_state_path,
pg_agent_config=config.pg_agent_config)
if config.title is not None:
wrapper_env.idsgame_env.idsgame_config.render_config.title = config.title
attacker = OpenAiPPOAgent(wrapper_env, config.pg_agent_config)
else:
raise AssertionError(
"Attacker train agent type not recognized: {}".format(
config.attacker_type))
attacker.train()
train_result = attacker.train_result
eval_result = attacker.eval_result
return train_result, eval_result
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=False,
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)
env_name = args.env_name
env = gym.make(env_name,
save_dir=default_output_dir() + "/results/data/" +
args.experiment_id)
attacker_agent = ReinforceAgent(env=env, config=pg_agent_config)
attacker_agent.train()
train_result = attacker_agent.train_result
eval_result = attacker_agent.eval_result