def day(self, actions, rewards):
"""
Run the day phase, that is execute target based on votes and reward accordingly or the voting
:param actions: dict, map id_ to vote
:param rewards: dict, maps agent id_ to curr reward
:return: updated rewards
"""
def execution(actions, rewards):
"""
To be called when is execution phase
:return:
"""
self.custom_metrics["suicide"] += suicide_num(actions)
# get the agent to be executed
target = most_frequent(actions)
# penalize for non divergent target
rewards = self.target_accord(target, rewards, actions)
# if target is alive
if self.status_map[target]:
# log
if Params.log_step == self.ep_step:
logger.debug(f"Player {target} ({self.role_map[target]}) has been executed")
# for every agent alive, [to be executed agent too]
for id_ in self.get_ids('all', alive=True):
# add/subtract penalty
if id_ == target:
rewards[id_] += self.penalties.get("death")
else:
rewards[id_] += self.penalties.get("execution")
# kill target
self.status_map[target] = 0
else:
# penalize agents for executing a dead one
for id_ in self.get_ids("all", alive=True):
rewards[id_] += self.penalties.get('execute_dead')
if Params.log_step == self.ep_step:
logger.debug(f"Players tried to execute dead agent {target}")
# update day
self.day_count += 1
return rewards
# call the appropriate method depending on the phase
if self.is_comm:
if Params.log_step == self.ep_step:
logger.debug("Day Time| Voting")
return rewards
else:
if Params.log_step == self.ep_step:
logger.debug("Day Time| Executing")
rewards = {id_: val + self.penalties.get('day') for id_, val in rewards.items()}
return execution(actions, rewards)
def execution(actions, rewards):
"""
To be called when is execution phase
:return:
"""
self.custom_metrics["suicide"] += suicide_num(actions)
# get the agent to be executed
target = most_frequent(actions)
# penalize for non divergent target
rewards = self.target_accord(target, rewards,
self.get_ids("all", alive=True))
# if target is alive
if self.status_map[target]:
# log
logger.debug(
f"Player {target} ({self.role_map[target]}) has been executed"
)
# for every agent alive, [to be executed agent too]
for id_ in [
elem for elem in rewards.keys()
if self.status_map[elem]
]:
# add/subtract penalty
if id_ == target:
rewards[id_] += self.penalties.get("death")
else:
rewards[id_] += self.penalties.get("execution")
# kill target
self.status_map[target] = 0
else:
# penalize agents for executing a dead one
for id_ in self.get_ids("all", alive=True):
rewards[id_] += self.penalties.get('execute_dead')
logger.debug(f"Players tried to execute dead agent {target}")
# increase the number of dead_man_execution in info
self.custom_metrics["dead_man_execution"] += 1
# update day
self.day_count += 1
return rewards
def step(self, actions):
"""Run one timestep of the environment's dynamics. When end of
episode is reached, you are responsible for calling `reset()`
to reset this environment's state.
Accepts an action and returns a tuple (observation, reward, done, info).
Args:
actions_dict (dict): a list of action provided by the agents
Returns:
observation (object): agent's observation of the current environment
reward (float) : amount of reward returned after previous action
done (bool): whether the episode has ended, in which case further step() calls will return undefined results
info (dict): contains auxiliary diagnostic information (helpful for debugging, and sometimes learning)
"""
# rewards start from zero
rewards = {id: 0 for id in self.get_ids("all", alive=False)}
# execute night action
if self.is_night:
logger.debug("Night Time")
rewards = self.night(actions, rewards)
self.is_night = not self.is_night
else: # else go with day
logger.debug("Day Time")
# penalize since a day has passed
rewards = {id: val + self.penalties.get('day') for id, val in rewards.items()}
rewards = self.day(actions, rewards)
self.is_night = not self.is_night
# update dones
dones, rewards = self.check_done(rewards)
obs = self.observe()
info = {id: self.infos for id in self.get_ids("all", alive=False)}
obs, rewards, dones, info = self.convert(obs, rewards, dones, info)
# if game over reset
if self.is_done:
self.infos["tot_days"] = self.day_count
dones["__all__"] = True
else:
dones["__all__"] = False
return obs, rewards, dones, info
def night(self, actions, rewards):
"""
Is night, time to perform actions!
During this phase, villagers action are not considered
:param actions: dict, map id_ to vote
:param rewards: dict, maps agent id_ to curr reward
:return: return updated rewards
"""
if self.is_comm:
logger.debug("Night Time| Voting")
else:
logger.debug("Night Time| Eating")
# execute wolf actions
rewards = self.wolf_action(actions, rewards)
return rewards
def kill(actions, rewards):
# upvote suicide info
self.custom_metrics["suicide"] += suicide_num(actions)
if not len(wolves_ids):
raise Exception(
"Game not done but wolves are dead, have reset been called?"
)
# get agent to be eaten
target = most_frequent(actions)
# penalize for different ids
rewards = self.target_accord(target, rewards, wolves_ids)
# if target is alive
if self.status_map[target]:
# kill him
self.status_map[target] = 0
# penalize dead player
rewards[target] += self.penalties.get("death")
# reward wolves
for id_ in wolves_ids:
rewards[id_] += self.penalties.get("kill")
logger.debug(
f"Wolves killed {target} ({self.role_map[target]})")
else:
logger.debug(f"Wolves tried to kill dead agent {target}")
# penalize the wolves for eating a dead player
for id_ in wolves_ids:
rewards[id_] += self.penalties.get('execute_dead')
# log it
self.custom_metrics["dead_man_kill"] += 1
if target in wolves_ids:
# penalize the agent for eating one of their kind
for id_ in wolves_ids:
rewards[id_] += self.penalties.get('kill_wolf')
# log it
self.custom_metrics["cannibalism"] += 1
return rewards
def wolf_action(self, actions, rewards):
"""
Perform wolf action, that is kill agent based on votes and reward
:param actions: dict, map id to vote
:param rewards: dict, maps agent id to curr reward
:return: updated rewards
"""
# get wolves ids
wolves_ids = self.get_ids(ww, alive=True)
# filter action to get only wolves
actions = {k: v for k, v in actions.items() if k in wolves_ids}
# upvote suicide info
self.infos["suicide"] += suicide_num(actions)
if not len(wolves_ids):
raise Exception("Game not done but wolves are dead")
# get choices by wolves
actions = [actions[id] for id in wolves_ids]
logger.debug(f"wolves votes :{actions}")
# get agent to be eaten
target = most_frequent(actions)
# if target is alive
if self.status_map[target]:
# kill him
self.status_map[target] = 0
# penalize dead player
rewards[target] += self.penalties.get("death")
# reward wolves
for id in wolves_ids:
rewards[id] += self.penalties.get("kill")
logger.debug(f"Wolves killed {target} ({self.role_map[target]})")
else:
logger.debug(f"Wolves tried to kill dead agent {target}")
# penalize the wolves for eating a dead player
for id in wolves_ids:
rewards[id] += self.penalties.get('execute_dead')
# log it
self.infos["dead_man_kill"] += 1
if target in wolves_ids:
# penalize the agent for eating one of their kind
for id in wolves_ids:
rewards[id] += self.penalties.get('kill_wolf')
# log it
self.infos["cannibalism"] += 1
return rewards
def day(self, actions, rewards):
"""
Run the day phase, that is execute target based on votes and reward accordingly
:param actions: dict, map id to vote
:param rewards: dict, maps agent id to curr reward
:return: updated rewards
"""
# update vote list
for idx in range(self.num_players):
# use -1 if agent is dead
self.votes[idx] = actions.get(idx, -1)
self.infos["suicide"] += suicide_num(actions)
# get the agent to be executed
target = most_frequent(actions)
logger.debug(f"Villagers votes {[elem for elem in actions.values()]}")
# if target is alive
if self.status_map[target]:
# log
logger.debug(f"Player {target} ({self.role_map[target]}) has been executed")
# kill target
self.status_map[target] = 0
# for every agent alive
for id in [elem for elem in rewards.keys() if self.status_map[elem]]:
# add/subtract penalty
if id == target:
rewards[id] += self.penalties.get("death")
else:
rewards[id] += self.penalties.get("execution")
else:
# penalize agents for executing a dead one
for id in self.get_ids("all", alive=True):
rewards[id] += self.penalties.get('execute_dead')
logger.debug(f"Players tried to execute dead agent {target}")
# increase the number of dead_man_execution in info
self.infos["dead_man_execution"] += 1
# update day
self.day_count += 1
return rewards