from typing import Union, List, Dict, Any
import numpy as np
from csle_common.dao.training.agent_type import AgentType
from csle_common.dao.training.player_type import PlayerType
from csle_common.dao.training.policy import Policy
from csle_common.dao.simulation_config.action import Action
from csle_common.dao.simulation_config.state import State
from csle_common.dao.training.policy_type import PolicyType
[docs]class AlphaVectorsPolicy(Policy):
"""
Object representing a policy based on alpha vectors for a POMDP (Sondik 1971)
"""
def __init__(self, player_type: PlayerType, actions: List[Action], alpha_vectors: List[Any],
transition_tensor: List[Any], reward_tensor: List[Any], states: List[State],
agent_type: AgentType, simulation_name: str, avg_R: float) -> None:
"""
Initializes the policy
:param actions: list of actions
:param states: list of states
:param player_type: the player type
:param alpha_vectors: the lookup table that defines the policy
:param value_function: the value function (optional)
:param simulation_name: the name of the simulation
:param avg_R: average reward obtained with the policy
:param transition_tensor: the transition tensor of the POMDP
:param reward_tensor: the reward tensor of the POMDP
"""
super(AlphaVectorsPolicy, self).__init__(agent_type=agent_type, player_type=player_type)
self.actions = actions
self.alpha_vectors = alpha_vectors
self.simulation_name = simulation_name
self.id = -1
self.avg_R = avg_R
self.transition_tensor = transition_tensor
self.reward_tensor = reward_tensor
self.states = states
self.policy_type = PolicyType.ALPHA_VECTORS
[docs] def action(self, o: List[Union[int, float]], deterministic: bool = True) -> int:
"""
Selects the next action
:param o: the belief
:param deterministic: boolean flag indicating whether the action selection should be deterministic
:return: the next action and its probability
"""
b = o
max_a_v = -np.inf
for a in self.actions:
v_a = 0
for s in self.states:
for s_prime in self.states:
transition_prob = (b[s.id] * self.reward_tensor[a.id][s.id] *
self.transition_tensor[a.id][s.id][s_prime.id])
max_alpha_v = -np.inf
for alpha in self.alpha_vectors:
v = np.dot(np.array(alpha), np.array(b[0:len(alpha)]))
if v > max_alpha_v:
max_alpha_v = v
v_a += max_alpha_v * transition_prob
if v_a > max_a_v:
max_a_v = v_a
max_a = a
try:
return max_a.id
except Exception:
return 0
[docs] def probability(self, o: List[Union[int, float]], a: int) -> float:
"""
Calculates the probability of taking a given action for a given observation
:param o: the input observation
:param a: the action
:return: p(a|o)
"""
return a == self.action(o=o)
[docs] @staticmethod
def from_dict(d: Dict[str, Any]) -> "AlphaVectorsPolicy":
"""
Converts a dict representation to an instance
:param d: the dict to convert
:return: the created instance
"""
dto = AlphaVectorsPolicy(actions=list(map(lambda x: Action.from_dict(x), d["actions"])),
player_type=d["player_type"], agent_type=d["agent_type"],
alpha_vectors=d["alpha_vectors"],
simulation_name=d["simulation_name"], avg_R=d["avg_R"],
transition_tensor=d["transition_tensor"], reward_tensor=d["reward_tensor"],
states=list(map(lambda x: State.from_dict(x), d["states"])))
if "id" in d:
dto.id = d["id"]
return dto
[docs] def to_dict(self) -> Dict[str, Any]:
"""
:return: A dict representation of the function
"""
d: Dict[str, Any] = {}
d["agent_type"] = self.agent_type
d["player_type"] = self.player_type
d["actions"] = list(map(lambda x: x.to_dict(), self.actions))
d["alpha_vectors"] = self.alpha_vectors
d["simulation_name"] = self.simulation_name
d["id"] = self.id
d["avg_R"] = self.avg_R
d["transition_tensor"] = self.transition_tensor
d["states"] = list(map(lambda x: x.to_dict(), self.states))
d["reward_tensor"] = self.reward_tensor
d["policy_type"] = self.policy_type
return d
[docs] def stage_policy(self, o: Union[List[Union[int, float]], int, float]) -> List[List[float]]:
"""
Gets the stage policy, i.e a |S|x|A| policy
:param o: the latest observation
:return: the |S|x|A| stage policy
"""
return self.alpha_vectors
def __str__(self) -> str:
"""
:return: a string representation of the policy
"""
return f"agent_type: {self.agent_type}, player_type: {self.player_type}, " \
f"actions: {list(map(lambda x: str(x), self.actions))}, alpha_vectors: {self.alpha_vectors}, " \
f"simulation_name: {self.simulation_name}, id: {self.id}, avg_R: {self.avg_R}," \
f"transition_tensor: {self.transition_tensor}, states: {self.states}, " \
f"reward_tensor: {self.reward_tensor}, policy_type: {self.policy_type}"
[docs] @staticmethod
def from_json_file(json_file_path: str) -> "AlphaVectorsPolicy":
"""
Reads a json file and converts it to a DTO
:param json_file_path: the json file path
:return: the converted DTO
"""
import io
import json
with io.open(json_file_path, 'r') as f:
json_str = f.read()
return AlphaVectorsPolicy.from_dict(json.loads(json_str))
[docs] def copy(self) -> "AlphaVectorsPolicy":
"""
:return: a copy of the DTO
"""
return AlphaVectorsPolicy.from_dict(self.to_dict())