steffen/marl-factory-grid
1
0
Fork 0
mirror of https://github.com/illiumst/marl-factory-grid.git synced 2025-09-15 07:23:59 +02:00
Code Issues Packages Projects Releases Wiki Activity

Accounting Zones for basefactory, mark them in level.txt by int, x for danger zones

Browse Source
This commit is contained in:
steffen-illium
2021-06-11 10:29:57 +02:00
parent b4ad112ea6
commit 99c01cd569
8 changed files with 119 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
environments/factory/base_factory.py
View File

@@ -7,20 +7,15 @@ import numpy as np
from gym import spaces
import yaml
from gym.wrappers import FrameStack
from environments import helpers as h
from environments.utility_classes import Actions, StateSlice, AgentState, MovementProperties
from environments.utility_classes import Actions, StateSlice, AgentState, MovementProperties, Zones
# noinspection PyAttributeOutsideInit
class BaseFactory(gym.Env):
def __setattr__(self, key, value):
if isinstance(value, dict):
super(BaseFactory, self).__setattr__(key, Namespace(**value))
else:
super(BaseFactory, self).__setattr__(key, value)
@property
def action_space(self):
return spaces.Discrete(self._actions.n)
@@ -40,11 +35,20 @@ class BaseFactory(gym.Env):
def movement_actions(self):
return self._actions.movement_actions
def __enter__(self):
return self if self.frames_to_stack == 0 else FrameStack(self, self.frames_to_stack)
def __exit__(self, exc_type, exc_val, exc_tb):
self.close()
def __init__(self, level_name='simple', n_agents=1, max_steps=int(5e2), pomdp_radius: Union[None, int] = 0,
movement_properties: MovementProperties = MovementProperties(),
combin_agent_slices_in_obs: bool = False,
combin_agent_slices_in_obs: bool = False, frames_to_stack=0,
omit_agent_slice_in_obs=False, **kwargs):
assert combin_agent_slices_in_obs != omit_agent_slice_in_obs, 'Both options are exclusive'
assert (combin_agent_slices_in_obs != omit_agent_slice_in_obs) or \
(not combin_agent_slices_in_obs and not omit_agent_slice_in_obs), \
'Both options are exclusive'
assert frames_to_stack != 1 and frames_to_stack >= 0, "'frames_to_stack' cannot be negative or 1."
self.movement_properties = movement_properties
self.level_name = level_name
@@ -54,17 +58,19 @@ class BaseFactory(gym.Env):
self.pomdp_radius = pomdp_radius
self.combin_agent_slices_in_obs = combin_agent_slices_in_obs
self.omit_agent_slice_in_obs = omit_agent_slice_in_obs
self.frames_to_stack = frames_to_stack
self.done_at_collision = False
_actions = Actions(self.movement_properties)
self._actions = _actions + self.additional_actions
self._level = h.one_hot_level(
h.parse_level(Path(__file__).parent / h.LEVELS_DIR / f'{self.level_name}.txt')
)
level_filepath = Path(__file__).parent / h.LEVELS_DIR / f'{self.level_name}.txt'
parsed_level = h.parse_level(level_filepath)
self._level = h.one_hot_level(parsed_level)
self._state_slices = StateSlice(n_agents)
if 'additional_slices' in kwargs:
self._state_slices.register_additional_items(kwargs.get('additional_slices'))
self._zones = Zones(parsed_level)
self.reset()
@property
@@ -259,7 +265,7 @@ class BaseFactory(gym.Env):
# d = {key: val._asdict() if hasattr(val, '_asdict') else val for key, val in self.__dict__.items()
d = {key: val for key, val in self.__dict__.items() if not key.startswith('_') and not key.startswith('__')}
filepath.parent.mkdir(parents=True, exist_ok=True)
super(BaseFactory, self).save_params()
with filepath.open('w') as f:
yaml.dump(d, f)
# pickle.dump(d, f, protocol=pickle.HIGHEST_PROTOCOL)

22
environments/factory/levels/rooms.txt
View File

@@ -1,13 +1,13 @@
###############
#------#------#
#---#--#------#
#--------#----#
#------#------#
#------#------#
###-#######-###
#----##-------#
#-----#----#--#
#-------------#
#-----#-------#
#-----#-------#
#333x33#444444#
#333#33#444444#
#333333xx#4444#
#333333#444444#
#333333#444444#
###x#######x###
#1111##2222222#
#11111#2222#22#
#11111x2222222#
#11111#2222222#
#11111#2222222#
###############

20
environments/factory/simple_factory.py
View File

@@ -8,7 +8,7 @@ from environments.factory.base_factory import BaseFactory
from environments import helpers as h
from environments.factory.renderer import Renderer, Entity
from environments.utility_classes import AgentState, MovementProperties
from environments.utility_classes import AgentState, MovementProperties, Register
DIRT_INDEX = -1
CLEAN_UP_ACTION = 'clean_up'
@@ -39,8 +39,8 @@ class SimpleFactory(BaseFactory):
self.dirt_properties = dirt_properties
self.verbose = verbose
self.max_dirt = 20
super(SimpleFactory, self).__init__(*args, additional_slices='dirt', **kwargs)
self._renderer = None # expensive - don't use it when not required !
super(SimpleFactory, self).__init__(*args, additional_slices='dirt', **kwargs)
def render(self):
@@ -79,7 +79,6 @@ class SimpleFactory(BaseFactory):
for x, y in free_for_dirt[:n_dirt_tiles]:
new_value = self._state[DIRT_INDEX, x, y] + self.dirt_properties.gain_amount
self._state[DIRT_INDEX, x, y] = max(new_value, self.dirt_properties.max_local_amount)
else:
pass
@@ -126,10 +125,11 @@ class SimpleFactory(BaseFactory):
return obs
def calculate_reward(self, agent_states: List[AgentState]) -> (int, dict):
# TODO: What reward to use?
info_dict = dict()
current_dirt_amount = self._state[DIRT_INDEX].sum()
dirty_tiles = np.argwhere(self._state[DIRT_INDEX] != h.IS_FREE_CELL).shape[0]
info_dict = dict()
info_dict.update(dirt_amount=current_dirt_amount)
info_dict.update(dirty_tile_count=dirty_tiles)
try:
# penalty = current_dirt_amount
@@ -156,14 +156,13 @@ class SimpleFactory(BaseFactory):
reward -= 0.01
self.print(f'Agent {agent_state.i} just tried to clean up some dirt '
f'at {agent_state.pos}, but was unsucsessfull.')
info_dict.update(failed_cleanup_attempt=1)
info_dict.update({f'agent_{agent_state.i}_failed_action': 1})
elif self._actions.is_moving_action(agent_state.action):
if agent_state.action_valid:
# info_dict.update(movement=1)
reward -= 0.00
else:
# info_dict.update(collision=1)
# self.print('collision')
reward -= 0.01
@@ -172,10 +171,9 @@ class SimpleFactory(BaseFactory):
reward -= 0.00
for entity in list_of_collisions:
entity = 'agent' if 'agent' in entity else entity
info_dict.update({f'agent_{agent_state.i}_vs_{entity}': 1})
info_dict.update(dirt_amount=current_dirt_amount)
info_dict.update(dirty_tile_count=dirty_tiles)
self.print(f"reward is {reward}")
# Potential based rewards ->
# track the last reward , minus the current reward = potential
@@ -191,8 +189,8 @@ if __name__ == '__main__':
move_props = MovementProperties(allow_diagonal_movement=True, allow_square_movement=True)
dirt_props = DirtProperties()
factory = SimpleFactory(movement_properties=move_props, dirt_properties=dirt_props, n_agents=2,
combin_agent_slices_in_obs=True, omit_agent_slice_in_obs=False)
factory = SimpleFactory(movement_properties=move_props, dirt_properties=dirt_props, n_agents=10,
combin_agent_slices_in_obs=True, omit_agent_slice_in_obs=False, level_name='rooms')
# dirt_props = DirtProperties()
# move_props = MovementProperties(allow_diagonal_movement=False, allow_no_op=False)