from typing import Union, List, Dict, Tuple, Optional, Any
import time
import gymnasium as gym
import os
import math
import numpy as np
from csle_common.dao.emulation_config.emulation_env_config import EmulationEnvConfig
from csle_common.dao.simulation_config.simulation_env_config import SimulationEnvConfig
from csle_common.dao.training.experiment_config import ExperimentConfig
from csle_common.dao.training.experiment_execution import ExperimentExecution
from csle_common.dao.training.experiment_result import ExperimentResult
from csle_common.dao.training.agent_type import AgentType
from csle_common.util.experiment_util import ExperimentUtil
from csle_common.dao.training.player_type import PlayerType
from csle_common.logging.log import Logger
from csle_common.metastore.metastore_facade import MetastoreFacade
from csle_common.dao.jobs.training_job_config import TrainingJobConfig
from csle_common.dao.training.mixed_ppo_policy import MixedPPOPolicy
from csle_common.dao.training.ppo_policy import PPOPolicy
from csle_agents.agents.ppo.ppo_agent import PPOAgent
from csle_common.dao.training.policy import Policy
from csle_common.dao.training.tabular_policy import TabularPolicy
import csle_common.constants.constants as constants
from csle_common.util.general_util import GeneralUtil
from csle_agents.agents.base.base_agent import BaseAgent
import csle_agents.constants.constants as agents_constants
import gym_csle_stopping_game.constants.constants as env_constants
from csle_common.dao.simulation_config.base_env import BaseEnv
[docs]class DFSPLocalPPOAgent(BaseAgent):
"""
RL Agent implementing the local DFSP algorithm
"""
def __init__(self, defender_simulation_env_config: SimulationEnvConfig,
attacker_simulation_env_config: SimulationEnvConfig,
emulation_env_config: Union[None, EmulationEnvConfig], experiment_config: ExperimentConfig,
training_job: Optional[TrainingJobConfig] = None):
"""
Initializes the local DFSP agent
:param attacker_simulation_env_config: the simulation env config of the attacker
:param defender_simulation_env_config: the simulation env config of the defender
:param emulation_env_config: the emulation env config
:param attacker_experiment_config: the experiment config
:param training_job: (optional) reuse an existing training job configuration
"""
super().__init__(simulation_env_config=defender_simulation_env_config,
emulation_env_config=emulation_env_config,
experiment_config=experiment_config)
self.root_output_dir = str(self.experiment_config.output_dir)
self.defender_experiment_config = self.get_defender_experiment_config()
self.attacker_experiment_config = self.get_attacker_experiment_config()
self.attacker_simulation_env_config = attacker_simulation_env_config
self.defender_simulation_env_config = defender_simulation_env_config
self.training_job = training_job
[docs] def train(self) -> ExperimentExecution:
"""
Performs the policy training for the given random seeds using the local DFSP algorithm
:return: the training metrics and the trained policies
"""
pid = os.getpid()
# Initialize result metrics
exp_result = ExperimentResult()
# Define which metrics to plot in the UI
exp_result.plot_metrics.append(agents_constants.COMMON.EXPLOITABILITY)
exp_result.plot_metrics.append(agents_constants.COMMON.RUNNING_AVERAGE_EXPLOITABILITY)
exp_result.plot_metrics.append(agents_constants.COMMON.AVERAGE_ATTACKER_RETURN)
exp_result.plot_metrics.append(agents_constants.COMMON.RUNNING_AVERAGE_ATTACKER_RETURN)
exp_result.plot_metrics.append(agents_constants.LOCAL_DFSP.RUNNING_AVERAGE_BEST_RESPONSE_ATTACKER_RETURN)
exp_result.plot_metrics.append(agents_constants.COMMON.AVERAGE_DEFENDER_RETURN)
exp_result.plot_metrics.append(agents_constants.COMMON.RUNNING_AVERAGE_DEFENDER_RETURN)
exp_result.plot_metrics.append(agents_constants.LOCAL_DFSP.RUNNING_AVERAGE_BEST_RESPONSE_DEFENDER_RETURN)
exp_result.plot_metrics.append(env_constants.ENV_METRICS.INTRUSION_LENGTH)
exp_result.plot_metrics.append(agents_constants.COMMON.RUNNING_AVERAGE_INTRUSION_LENGTH)
exp_result.plot_metrics.append(env_constants.ENV_METRICS.INTRUSION_START)
exp_result.plot_metrics.append(agents_constants.COMMON.RUNNING_AVERAGE_INTRUSION_START)
exp_result.plot_metrics.append(env_constants.ENV_METRICS.TIME_HORIZON)
exp_result.plot_metrics.append(agents_constants.COMMON.RUNNING_AVERAGE_TIME_HORIZON)
exp_result.plot_metrics.append(env_constants.ENV_METRICS.AVERAGE_UPPER_BOUND_RETURN)
descr = f"Approximating a Nash equilibrium with the local DFSP algorithm using " \
f"simulations: {self.defender_simulation_env_config.name} " \
f"and {self.attacker_simulation_env_config.name}"
for seed in self.experiment_config.random_seeds:
exp_result.all_metrics[seed] = {}
exp_result.all_metrics[seed][agents_constants.COMMON.AVERAGE_DEFENDER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_DEFENDER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.COMMON.AVERAGE_ATTACKER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_ATTACKER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.AVERAGE_BEST_RESPONSE_DEFENDER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.RUNNING_AVERAGE_BEST_RESPONSE_DEFENDER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.AVERAGE_BEST_RESPONSE_ATTACKER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.RUNNING_AVERAGE_BEST_RESPONSE_ATTACKER_RETURN] = []
exp_result.all_metrics[seed][agents_constants.COMMON.EXPLOITABILITY] = []
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_EXPLOITABILITY] = []
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_INTRUSION_START] = []
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_TIME_HORIZON] = []
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_INTRUSION_LENGTH] = []
exp_result.all_metrics[seed][env_constants.ENV_METRICS.INTRUSION_START] = []
exp_result.all_metrics[seed][env_constants.ENV_METRICS.INTRUSION_LENGTH] = []
exp_result.all_metrics[seed][env_constants.ENV_METRICS.TIME_HORIZON] = []
exp_result.all_metrics[seed][env_constants.ENV_METRICS.AVERAGE_UPPER_BOUND_RETURN] = []
if self.training_job is None:
emulation_name = ""
if self.emulation_env_config is not None:
emulation_name = self.emulation_env_config.name
self.training_job = TrainingJobConfig(
simulation_env_name=self.simulation_env_config.name, experiment_config=self.experiment_config,
experiment_result=exp_result, progress_percentage=0, pid=pid,
emulation_env_name=emulation_name, simulation_traces=[],
num_cached_traces=agents_constants.COMMON.NUM_CACHED_SIMULATION_TRACES,
log_file_path=Logger.__call__().get_log_file_path(), descr=descr,
physical_host_ip=GeneralUtil.get_host_ip())
training_job_id = MetastoreFacade.save_training_job(training_job=self.training_job)
self.training_job.id = training_job_id
else:
self.training_job.pid = pid
self.training_job.progress_percentage = 0
self.training_job.experiment_result = exp_result
MetastoreFacade.update_training_job(training_job=self.training_job, id=self.training_job.id)
config = self.simulation_env_config.simulation_env_input_config
env = gym.make(self.simulation_env_config.gym_env_name, config=config)
for seed in self.experiment_config.random_seeds:
ExperimentUtil.set_seed(seed)
exp_result = self.local_dfsp(exp_result=exp_result, seed=seed, env=env, training_job=self.training_job,
random_seeds=self.experiment_config.random_seeds)
self.training_job = MetastoreFacade.get_training_job_config(id=self.training_job.id)
# Calculate average and std metrics
exp_result.avg_metrics = {}
exp_result.std_metrics = {}
for metric in exp_result.all_metrics[self.experiment_config.random_seeds[0]].keys():
confidence = 0.95
value_vectors = []
for seed in self.experiment_config.random_seeds:
value_vectors.append(exp_result.all_metrics[seed][metric])
avg_metrics = []
std_metrics = []
for i in range(len(value_vectors[0])):
seed_values = []
for seed_idx in range(len(self.experiment_config.random_seeds)):
seed_values.append(value_vectors[seed_idx][i])
avg_metrics.append(ExperimentUtil.mean_confidence_interval(data=seed_values, confidence=confidence)[0])
std_metrics.append(ExperimentUtil.mean_confidence_interval(data=seed_values, confidence=confidence)[1])
exp_result.avg_metrics[metric] = avg_metrics
exp_result.std_metrics[metric] = std_metrics
ts = time.time()
emulation_name = ""
if self.emulation_env_config is not None:
emulation_name = self.emulation_env_config.name
simulation_name = self.simulation_env_config.name
exp_execution = ExperimentExecution(result=exp_result, config=self.experiment_config, timestamp=ts,
emulation_name=emulation_name, simulation_name=simulation_name,
descr=descr, log_file_path=self.training_job.log_file_path)
traces = []
if isinstance(env, BaseEnv):
traces = env.get_traces()
if len(traces) > 0:
MetastoreFacade.save_simulation_trace(traces[-1])
MetastoreFacade.remove_training_job(self.training_job)
return exp_execution
[docs] def local_dfsp(self, exp_result: ExperimentResult, seed: int, env: BaseEnv,
training_job: TrainingJobConfig, random_seeds: List[int]) -> ExperimentResult:
"""
Implements the local DFSP training logic
:param exp_result: the experiment result
:param seed: the seed for the experiments
:param env: the environment for the experiment
:param training_job: the training job
:param random_seeds: the random seeds for the experiment
:return: None
"""
# Initialize policies
defender_strategy = MixedPPOPolicy(
simulation_name=self.defender_simulation_env_config.name,
states=self.defender_simulation_env_config.state_space_config.states,
player_type=PlayerType.DEFENDER,
actions=self.defender_simulation_env_config.joint_action_space_config.action_spaces[
self.defender_experiment_config.player_idx].actions,
experiment_config=self.defender_experiment_config, avg_R=-1)
defender_strategy.states = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_D
defender_strategy.actions = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A1
attacker_strategy = MixedPPOPolicy(
simulation_name=self.attacker_simulation_env_config.name,
states=self.attacker_simulation_env_config.state_space_config.states,
player_type=PlayerType.ATTACKER,
actions=self.attacker_simulation_env_config.joint_action_space_config.action_spaces[
self.attacker_experiment_config.player_idx].actions,
experiment_config=self.attacker_experiment_config, avg_R=-1)
attacker_strategy.states = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_A
attacker_strategy.actions = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A2
for i in range(self.experiment_config.hparams[agents_constants.LOCAL_DFSP.N_2].value):
# Compute best responses
br_seed = np.random.randint(0, 100)
attacker_br, attacker_val = self.attacker_best_response(
seed=br_seed, defender_strategy=defender_strategy, attacker_strategy=attacker_strategy)
defender_br, defender_val = self.defender_best_response(
seed=br_seed, attacker_strategy=attacker_strategy, defender_strategy=defender_strategy)
# Update strategies
defender_br.states = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_D
defender_br.actions = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A1
attacker_br.states = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_A
attacker_br.actions = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A2
# Update empirical strategies
attacker_strategy.ppo_policies.append(attacker_br)
defender_strategy.ppo_policies.append(defender_br)
# Evaluate best response strategies against empirical strategies
attacker_metrics = self.evaluate_attacker_policy(
defender_strategy=self.attacker_simulation_env_config.simulation_env_input_config.defender_strategy,
attacker_strategy=attacker_br)
# defender_metrics = self.evaluate_defender_policy(
# attacker_strategy=attacker_strategy, defender_strategy=defender_br)
defender_metrics = self.evaluate_attacker_policy(
attacker_strategy=self.defender_simulation_env_config.simulation_env_input_config.attacker_strategy,
defender_strategy=defender_br)
# Evaluate empirical against empirical
strategy_profile_metrics = self.evaluate_strategy_profile(
defender_strategy=defender_strategy, attacker_strategy=attacker_strategy)
# Update envs for the next BR iteration
self.attacker_simulation_env_config.simulation_env_input_config.defender_strategy = defender_strategy
self.attacker_simulation_env_config.simulation_env_input_config.attacker_strategy = attacker_strategy
self.defender_simulation_env_config.simulation_env_input_config.defender_strategy = defender_strategy
self.defender_simulation_env_config.simulation_env_input_config.attacker_strategy = attacker_strategy
# Compute exploitability
attacker_val = round(attacker_metrics[env_constants.ENV_METRICS.RETURN], 3)
defender_val = -round(defender_metrics[env_constants.ENV_METRICS.RETURN], 3)
attacker_val = max(attacker_val, -defender_val)
defender_val = max(defender_val, -attacker_val)
val = -round(strategy_profile_metrics[env_constants.ENV_METRICS.RETURN], 3)
val_attacker_exp = attacker_val
val_defender_exp = defender_val
# Don't log the first iteration which is just initializing the policies
if i == 0:
continue
# Log rewards
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.AVERAGE_BEST_RESPONSE_ATTACKER_RETURN].append(
val_attacker_exp)
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.AVERAGE_BEST_RESPONSE_DEFENDER_RETURN].append(
val_defender_exp)
exp_result.all_metrics[seed][
agents_constants.LOCAL_DFSP.RUNNING_AVERAGE_BEST_RESPONSE_ATTACKER_RETURN].append(
ExperimentUtil.running_average(
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.AVERAGE_BEST_RESPONSE_ATTACKER_RETURN],
self.experiment_config.hparams[agents_constants.COMMON.RUNNING_AVERAGE].value))
exp_result.all_metrics[seed][
agents_constants.LOCAL_DFSP.RUNNING_AVERAGE_BEST_RESPONSE_DEFENDER_RETURN].append(
ExperimentUtil.running_average(
exp_result.all_metrics[seed][agents_constants.LOCAL_DFSP.AVERAGE_BEST_RESPONSE_DEFENDER_RETURN],
self.experiment_config.hparams[agents_constants.COMMON.RUNNING_AVERAGE].value))
exp_result.all_metrics[seed][agents_constants.COMMON.AVERAGE_ATTACKER_RETURN].append(val)
exp_result.all_metrics[seed][agents_constants.COMMON.AVERAGE_DEFENDER_RETURN].append(-val)
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_ATTACKER_RETURN].append(
ExperimentUtil.running_average(
exp_result.all_metrics[seed][agents_constants.COMMON.AVERAGE_ATTACKER_RETURN],
self.experiment_config.hparams[agents_constants.COMMON.RUNNING_AVERAGE].value))
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_DEFENDER_RETURN].append(
ExperimentUtil.running_average(
exp_result.all_metrics[seed][agents_constants.COMMON.AVERAGE_DEFENDER_RETURN],
self.experiment_config.hparams[agents_constants.COMMON.RUNNING_AVERAGE].value))
exp_result.all_metrics[seed][env_constants.ENV_METRICS.TIME_HORIZON].append(
round(strategy_profile_metrics[env_constants.ENV_METRICS.TIME_HORIZON], 3))
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_TIME_HORIZON].append(
ExperimentUtil.running_average(
exp_result.all_metrics[seed][env_constants.ENV_METRICS.TIME_HORIZON],
self.experiment_config.hparams[agents_constants.COMMON.RUNNING_AVERAGE].value))
# Log baseline returns
exp_result.all_metrics[seed][env_constants.ENV_METRICS.AVERAGE_UPPER_BOUND_RETURN].append(
round(strategy_profile_metrics[env_constants.ENV_METRICS.AVERAGE_UPPER_BOUND_RETURN], 3))
# Compute and log exploitability
exp = DFSPLocalPPOAgent.exploitability(attacker_val=val_attacker_exp, defender_val=val_defender_exp)
exp_result.all_metrics[seed][agents_constants.COMMON.EXPLOITABILITY].append(exp)
running_avg_exp = ExperimentUtil.running_average(
exp_result.all_metrics[seed][agents_constants.COMMON.EXPLOITABILITY],
self.experiment_config.hparams[agents_constants.COMMON.RUNNING_AVERAGE].value)
exp_result.all_metrics[seed][agents_constants.COMMON.RUNNING_AVERAGE_EXPLOITABILITY].append(running_avg_exp)
# Logging the progress
if i % self.experiment_config.log_every == 0:
Logger.__call__().get_logger().info(
f"[Local DFSP] i: {i}, Exp: {exp}, "
f"Exp_avg_{self.experiment_config.hparams[agents_constants.COMMON.RUNNING_AVERAGE].value}: "
f"{running_avg_exp}, game_val: {val} "
f"opt_val:{exp_result.all_metrics[seed][env_constants.ENV_METRICS.AVERAGE_UPPER_BOUND_RETURN][-1]},"
f" Defender val:{defender_val}, Attacker val:{attacker_val}")
# Update training job
total_iterations = len(random_seeds) * self.experiment_config.hparams[
agents_constants.LOCAL_DFSP.N_2].value
iterations_done = ((random_seeds.index(seed)) *
self.experiment_config.hparams[agents_constants.LOCAL_DFSP.N_2].value + i)
progress = round(iterations_done / total_iterations, 2)
training_job.progress_percentage = progress
MetastoreFacade.update_training_job(training_job=training_job, id=training_job.id)
return exp_result
[docs] def evaluate_defender_policy(self, defender_strategy: Policy,
attacker_strategy: Policy) -> Dict[str, Union[float, int]]:
"""
Monte-Carlo evaluation of the game value of a given defender policy against the average attacker strategy
:param defender_thresholds: the defender strategy to evaluate
:param attacker_strategy: the average attacker strategy
:return: the average reward
"""
attacker_strategy.states = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_A
attacker_strategy.actions = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A2
defender_strategy.states = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_D
defender_strategy.actions = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A1
self.defender_simulation_env_config.simulation_env_input_config.attacker_strategy = attacker_strategy
self.defender_simulation_env_config.simulation_env_input_config.defender_strategy = defender_strategy
env = gym.make(self.defender_simulation_env_config.gym_env_name,
config=self.defender_simulation_env_config.simulation_env_input_config)
return self._eval_env(
env=env, policy=defender_strategy,
num_iterations=self.experiment_config.hparams[
agents_constants.LOCAL_DFSP.BEST_RESPONSE_EVALUATION_ITERATIONS].value)
[docs] def evaluate_strategy_profile(self, defender_strategy: MixedPPOPolicy,
attacker_strategy: MixedPPOPolicy) -> Dict[str, Union[float, int]]:
"""
Monte-Carlo evaluation of the game value following a given strategy profile
:param defender_strategy: the average defender strategy
:param attacker_strategy: the average attacker strategy
:return: the average reward
"""
defender_strategy.states = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_D
defender_strategy.actions = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A1
attacker_strategy.states = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_A
attacker_strategy.actions = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A2
self.attacker_simulation_env_config.simulation_env_input_config.defender_strategy = defender_strategy
self.attacker_simulation_env_config.simulation_env_input_config.attacker_strategy = attacker_strategy
env = gym.make(self.attacker_simulation_env_config.gym_env_name,
config=self.attacker_simulation_env_config.simulation_env_input_config)
return self._eval_env(
env=env, policy=attacker_strategy,
num_iterations=self.experiment_config.hparams[
agents_constants.LOCAL_DFSP.EQUILIBRIUM_STRATEGIES_EVALUATION_ITERATIONS].value)
[docs] def evaluate_attacker_policy(self, defender_strategy: Policy,
attacker_strategy: Policy) -> Dict[str, Union[float, int]]:
"""
Monte-Carlo evaluation of the game value of a given attacker strategy against the average defender strategy
:param defender_strategy: the average defender strategy
:param attacker_strategy: the attacker strategy to evaluate
:return: the average reward
"""
defender_strategy.states = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_D
defender_strategy.actions = \
self.defender_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A1
attacker_strategy.states = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.S_A
attacker_strategy.actions = \
self.attacker_simulation_env_config.simulation_env_input_config.local_intrusion_response_game_config.A2
self.attacker_simulation_env_config.simulation_env_input_config.defender_strategy = defender_strategy
self.attacker_simulation_env_config.simulation_env_input_config.attacker_strategy = attacker_strategy
env = gym.make(self.attacker_simulation_env_config.gym_env_name,
config=self.attacker_simulation_env_config.simulation_env_input_config)
return self._eval_env(
env=env, policy=attacker_strategy,
num_iterations=self.experiment_config.hparams[
agents_constants.LOCAL_DFSP.BEST_RESPONSE_EVALUATION_ITERATIONS].value)
[docs] def defender_best_response(self, seed: int, defender_strategy: MixedPPOPolicy,
attacker_strategy: MixedPPOPolicy) -> Tuple[PPOPolicy, float]:
"""
Learns a best response for the defender against a given attacker strategy
:param seed: the random seed
:param defender_strategy: the defender strategy
:param attacker_strategy: the attacker strategy
:return: the learned best response strategy and the average return
"""
self.defender_experiment_config.random_seeds = [seed]
self.defender_experiment_config.output_dir = str(self.root_output_dir)
self.defender_experiment_config.agent_type = AgentType.PPO
agent = PPOAgent(emulation_env_config=self.emulation_env_config,
simulation_env_config=self.defender_simulation_env_config,
experiment_config=self.defender_experiment_config, save_to_metastore=False)
Logger.__call__().get_logger().info(f"[Local DFSP] Starting training of the defender's best response "
f"against attacker strategy: {attacker_strategy}")
experiment_execution = agent.train()
policy: PPOPolicy = experiment_execution.result.policies[seed]
val = experiment_execution.result.avg_metrics[agents_constants.COMMON.RUNNING_AVERAGE_RETURN][-1]
return policy, val
def _eval_env(self, env: BaseEnv, policy: Policy, num_iterations: int) -> Dict[str, Union[float, int]]:
"""
:param env: the environment to use for evaluation
:param policy: the policy to evaluate
:param num_iterations: number of iterations to evaluate
:return: the average reward
"""
metrics: Dict[str, Any] = {}
for j in range(num_iterations):
done = False
o, _ = env.reset()
J = 0
t = 1
while not done and t <= self.experiment_config.hparams[agents_constants.COMMON.MAX_ENV_STEPS].value:
if isinstance(policy, TabularPolicy):
o = int(o[0])
a = policy.action(o=o)
o, r, done, _, info = env.step(a)
J += r
t += 1
metrics = self.update_metrics(metrics=metrics, info=info)
avg_metrics = self.compute_avg_metrics(metrics=metrics)
return avg_metrics
[docs] @staticmethod
def update_metrics(metrics: Dict[str, List[Union[float, int]]], info: Dict[str, Union[float, int]]) \
-> Dict[str, List[Union[float, int]]]:
"""
Update a dict with aggregated metrics using new information from the environment
:param metrics: the dict with the aggregated metrics
:param info: the new information
:return: the updated dict
"""
for k, v in info.items():
try:
if k in metrics:
metrics[k].append(round(v, 3))
else:
metrics[k] = [v]
except Exception:
pass
return metrics
[docs] @staticmethod
def compute_avg_metrics(metrics: Dict[str, List[Union[float, int]]]) -> Dict[str, Union[float, int]]:
"""
Computes the average metrics of a dict with aggregated metrics
:param metrics: the dict with the aggregated metrics
:return: the average metrics
"""
avg_metrics = {}
for k, v in metrics.items():
try:
avg = round(sum(v) / len(v), 2)
avg_metrics[k] = avg
except Exception:
pass
return avg_metrics
[docs] def attacker_best_response(self, seed: int, defender_strategy: MixedPPOPolicy, attacker_strategy: MixedPPOPolicy) \
-> Tuple[PPOPolicy, float]:
"""
Learns a best response strategy for the attacker against a given defender strategy
:param seed: the random seed
:param defender_strategy: the defender strategy
:param attacker_strategy: the attacker strategy
:return: the learned best response strategy and the average return
"""
self.attacker_experiment_config.random_seeds = [seed]
self.attacker_experiment_config.output_dir = str(self.root_output_dir)
self.attacker_experiment_config.agent_type = AgentType.PPO
agent = PPOAgent(emulation_env_config=self.emulation_env_config,
simulation_env_config=self.attacker_simulation_env_config,
experiment_config=self.attacker_experiment_config, save_to_metastore=False)
Logger.__call__().get_logger().info(f"[Local DFSP] Starting training of the attacker's best response "
f"against defender strategy: {defender_strategy}")
experiment_execution = agent.train()
policy: PPOPolicy = experiment_execution.result.policies[seed]
val = experiment_execution.result.avg_metrics[agents_constants.COMMON.RUNNING_AVERAGE_RETURN][-1]
return policy, val
[docs] def hparam_names(self) -> List[str]:
"""
:return: a list with the hyperparameter names
"""
return [
constants.NEURAL_NETWORKS.NUM_NEURONS_PER_HIDDEN_LAYER,
constants.NEURAL_NETWORKS.NUM_HIDDEN_LAYERS,
agents_constants.PPO.STEPS_BETWEEN_UPDATES,
agents_constants.COMMON.LEARNING_RATE, agents_constants.COMMON.BATCH_SIZE,
agents_constants.COMMON.GAMMA, agents_constants.PPO.GAE_LAMBDA, agents_constants.PPO.CLIP_RANGE,
agents_constants.PPO.CLIP_RANGE_VF, agents_constants.PPO.ENT_COEF,
agents_constants.PPO.VF_COEF, agents_constants.PPO.MAX_GRAD_NORM, agents_constants.PPO.TARGET_KL,
agents_constants.COMMON.NUM_TRAINING_TIMESTEPS, agents_constants.COMMON.EVAL_EVERY,
agents_constants.COMMON.EVAL_BATCH_SIZE, constants.NEURAL_NETWORKS.DEVICE,
agents_constants.COMMON.EVAL_BATCH_SIZE,
agents_constants.LOCAL_DFSP.N_2,
agents_constants.COMMON.CONFIDENCE_INTERVAL, agents_constants.COMMON.RUNNING_AVERAGE,
agents_constants.LOCAL_DFSP.BEST_RESPONSE_EVALUATION_ITERATIONS,
agents_constants.LOCAL_DFSP.EQUILIBRIUM_STRATEGIES_EVALUATION_ITERATIONS]
[docs] @staticmethod
def exploitability(attacker_val: float, defender_val: float) -> float:
"""
Computes the exploitability metric given the value of the attacker when following a best response
against the current defender strategy and the value of the defender when following a best response against
the current attacker strategy.
:param attacker_val: the value of the attacker when following a best response against
the current defender strategy
:param defender_val: the value of the defender when following a best response against the
current attacker strategy
:return: the exploitability
"""
return round(math.fabs(attacker_val + defender_val), 2)
[docs] def get_defender_experiment_config(self) -> ExperimentConfig:
"""
:return: the experiment configuration for learning a best response of the defender
"""
hparams = {
constants.NEURAL_NETWORKS.NUM_NEURONS_PER_HIDDEN_LAYER: self.experiment_config.hparams[
constants.NEURAL_NETWORKS.NUM_NEURONS_PER_HIDDEN_LAYER],
constants.NEURAL_NETWORKS.NUM_HIDDEN_LAYERS: self.experiment_config.hparams[
constants.NEURAL_NETWORKS.NUM_HIDDEN_LAYERS],
agents_constants.PPO.STEPS_BETWEEN_UPDATES: self.experiment_config.hparams[
agents_constants.PPO.STEPS_BETWEEN_UPDATES],
agents_constants.COMMON.BATCH_SIZE: self.experiment_config.hparams[agents_constants.COMMON.BATCH_SIZE],
agents_constants.COMMON.LEARNING_RATE: self.experiment_config.hparams[
agents_constants.COMMON.LEARNING_RATE],
constants.NEURAL_NETWORKS.DEVICE: self.experiment_config.hparams[constants.NEURAL_NETWORKS.DEVICE],
agents_constants.COMMON.NUM_PARALLEL_ENVS: self.experiment_config.hparams[
agents_constants.COMMON.NUM_PARALLEL_ENVS],
agents_constants.COMMON.GAMMA: self.experiment_config.hparams[agents_constants.COMMON.GAMMA],
agents_constants.PPO.GAE_LAMBDA: self.experiment_config.hparams[agents_constants.PPO.GAE_LAMBDA],
agents_constants.PPO.CLIP_RANGE: self.experiment_config.hparams[agents_constants.PPO.CLIP_RANGE],
agents_constants.PPO.CLIP_RANGE_VF: self.experiment_config.hparams[agents_constants.PPO.CLIP_RANGE_VF],
agents_constants.PPO.ENT_COEF: self.experiment_config.hparams[agents_constants.PPO.ENT_COEF],
agents_constants.PPO.VF_COEF: self.experiment_config.hparams[agents_constants.PPO.VF_COEF],
agents_constants.PPO.MAX_GRAD_NORM: self.experiment_config.hparams[agents_constants.PPO.MAX_GRAD_NORM],
agents_constants.PPO.TARGET_KL: self.experiment_config.hparams[agents_constants.PPO.TARGET_KL],
agents_constants.COMMON.NUM_TRAINING_TIMESTEPS: self.experiment_config.hparams[
agents_constants.COMMON.NUM_TRAINING_TIMESTEPS],
agents_constants.COMMON.EVAL_EVERY: self.experiment_config.hparams[agents_constants.COMMON.EVAL_EVERY],
agents_constants.COMMON.EVAL_BATCH_SIZE: self.experiment_config.hparams[
agents_constants.COMMON.EVAL_BATCH_SIZE],
agents_constants.COMMON.SAVE_EVERY: self.experiment_config.hparams[agents_constants.COMMON.SAVE_EVERY],
agents_constants.COMMON.CONFIDENCE_INTERVAL: self.experiment_config.hparams[
agents_constants.COMMON.CONFIDENCE_INTERVAL],
agents_constants.COMMON.MAX_ENV_STEPS: self.experiment_config.hparams[
agents_constants.COMMON.MAX_ENV_STEPS],
agents_constants.COMMON.RUNNING_AVERAGE: self.experiment_config.hparams[
agents_constants.COMMON.RUNNING_AVERAGE]
}
return ExperimentConfig(
output_dir=str(self.root_output_dir),
title="Learning a best response of the defender as part of local DFSP",
random_seeds=[], agent_type=AgentType.PPO,
log_every=self.experiment_config.br_log_every,
hparams=hparams,
player_type=PlayerType.DEFENDER, player_idx=0
)
[docs] def get_attacker_experiment_config(self) -> ExperimentConfig:
"""
:return: the experiment configuration for learning a best response of the attacker
"""
hparams = {
constants.NEURAL_NETWORKS.NUM_NEURONS_PER_HIDDEN_LAYER: self.experiment_config.hparams[
constants.NEURAL_NETWORKS.NUM_NEURONS_PER_HIDDEN_LAYER],
constants.NEURAL_NETWORKS.NUM_HIDDEN_LAYERS: self.experiment_config.hparams[
constants.NEURAL_NETWORKS.NUM_HIDDEN_LAYERS],
agents_constants.PPO.STEPS_BETWEEN_UPDATES: self.experiment_config.hparams[
agents_constants.PPO.STEPS_BETWEEN_UPDATES],
agents_constants.COMMON.BATCH_SIZE: self.experiment_config.hparams[agents_constants.COMMON.BATCH_SIZE],
agents_constants.COMMON.LEARNING_RATE: self.experiment_config.hparams[
agents_constants.COMMON.LEARNING_RATE],
constants.NEURAL_NETWORKS.DEVICE: self.experiment_config.hparams[constants.NEURAL_NETWORKS.DEVICE],
agents_constants.COMMON.NUM_PARALLEL_ENVS: self.experiment_config.hparams[
agents_constants.COMMON.NUM_PARALLEL_ENVS],
agents_constants.COMMON.GAMMA: self.experiment_config.hparams[agents_constants.COMMON.GAMMA],
agents_constants.PPO.GAE_LAMBDA: self.experiment_config.hparams[agents_constants.PPO.GAE_LAMBDA],
agents_constants.PPO.CLIP_RANGE: self.experiment_config.hparams[agents_constants.PPO.CLIP_RANGE],
agents_constants.PPO.CLIP_RANGE_VF: self.experiment_config.hparams[agents_constants.PPO.CLIP_RANGE_VF],
agents_constants.PPO.ENT_COEF: self.experiment_config.hparams[agents_constants.PPO.ENT_COEF],
agents_constants.PPO.VF_COEF: self.experiment_config.hparams[agents_constants.PPO.VF_COEF],
agents_constants.PPO.MAX_GRAD_NORM: self.experiment_config.hparams[agents_constants.PPO.MAX_GRAD_NORM],
agents_constants.PPO.TARGET_KL: self.experiment_config.hparams[agents_constants.PPO.TARGET_KL],
agents_constants.COMMON.NUM_TRAINING_TIMESTEPS: self.experiment_config.hparams[
agents_constants.COMMON.NUM_TRAINING_TIMESTEPS],
agents_constants.COMMON.EVAL_EVERY: self.experiment_config.hparams[agents_constants.COMMON.EVAL_EVERY],
agents_constants.COMMON.EVAL_BATCH_SIZE: self.experiment_config.hparams[
agents_constants.COMMON.EVAL_BATCH_SIZE],
agents_constants.COMMON.SAVE_EVERY: self.experiment_config.hparams[agents_constants.COMMON.SAVE_EVERY],
agents_constants.COMMON.CONFIDENCE_INTERVAL: self.experiment_config.hparams[
agents_constants.COMMON.CONFIDENCE_INTERVAL],
agents_constants.COMMON.MAX_ENV_STEPS: self.experiment_config.hparams[
agents_constants.COMMON.MAX_ENV_STEPS],
agents_constants.COMMON.RUNNING_AVERAGE: self.experiment_config.hparams[
agents_constants.COMMON.RUNNING_AVERAGE]
}
return ExperimentConfig(
output_dir=str(self.root_output_dir),
title="Learning a best response of the attacker as part of local DFSP",
random_seeds=[], agent_type=AgentType.T_SPSA,
log_every=self.experiment_config.br_log_every,
hparams=hparams,
player_type=PlayerType.ATTACKER, player_idx=1
)
[docs] @staticmethod
def round_vec(vec) -> List[float]:
"""
Rounds a vector to 3 decimals
:param vec: the vector to round
:return: the rounded vector
"""
return list(map(lambda x: round(x, 3), vec))
[docs] @staticmethod
def running_average(x: List[float], N: int) -> List[float]:
"""
Calculates the running average of the last N elements of vector x
:param x: the vector
:param N: the number of elements to use for average calculation
:return: the running average vector
"""
if len(x) >= N:
y = np.copy(x)
y[N - 1:] = np.convolve(x, np.ones((N,)) / N, mode='valid')
else:
N = len(x)
y = np.copy(x)
y[N - 1:] = np.convolve(x, np.ones((N,)) / N, mode='valid')
return list(y.tolist())