def _next_pos(self, pos, direction):
direction = Moves.get_coord(direction)
if pos.x == 0 and pos.y == self.board['_passage_y'] and\
direction == Moves.EAST:
next_pos = Coord(self.grid.x_size - 1, pos.y)
elif pos.x == self.grid.x_size - 1 and\
pos.y == self.board['_passage_y'] and direction == Moves.WEST:
next_pos = Coord(0, pos.y)
else:
next_pos = pos + direction
if self.grid.is_inside(next_pos) and self.is_passable(next_pos):
return next_pos
else:
return Coord(-1, -1)
class Compass(Enum):
North = Coord(0, 1)
East = Coord(1, 0)
South = Coord(0, -1)
West = Coord(-1, 0)
Null = Coord(0, 0)
NorthEast = Coord(1, 1)
SouthEast = Coord(1, -1)
SouthWest = Coord(-1, -1)
NorthWest = Coord(-1, 1)
@staticmethod
def get_coord(idx):
return list(Compass)[idx].value
def _get_init_state(self):
self.state = PocState()
self.state.agent_pos = Coord(*self.board["_poc_home"])
ghost_home = Coord(*self.board["_ghost_home"])
for g in range(self.board["_num_ghosts"]):
pos = Coord(ghost_home.x + g % 2, ghost_home.y + g // 2)
self.state.ghosts.append(Ghost(pos, direction=-1))
self.state.food_pos = np.random.binomial(1,
config["_food_prob"],
size=self.grid.n_tiles)
# only make free space food
idx = (self.board["_maze"] > 0) &\
(self.state.food_pos.reshape(self.board["_maze"].shape) > 0)
self.board["_maze"][idx] = 4
self.state.power_step = 0
return self.state
def _smell_food(self, smell_range=1):
for x in range(-smell_range, smell_range + 1):
for y in range(-smell_range, smell_range + 1):
smell_pos = Coord(x, y)
idx = self.grid.get_index(self.state.agent_pos + smell_pos)
if self.grid.is_inside(self.state.agent_pos + smell_pos) and\
self.state.food_pos[idx]:
return True
return False
def _decode_state(self, state):
poc_state = PocState(Coord(*self.grid.get_coord(state[0])))
ghosts = np.split(state[1:self.board["_num_ghosts"] * 3], 1)
for g in ghosts:
poc_state.ghosts.append(
Ghost(pos=self.grid.get_coord(g[0]), direction=g[1]))
poc_state.power_step = state[-1]
poc_state.food_pos = state[self.board["_num_ghosts"] * 3:-1].tolist()
return poc_state
def _get_init_state(self):
# create walls
# for tile in self.grid:
# value = config["maze"][tile.key[0]]
# self.grid.set_value(value, coord=tile.key)
self.state = PocState()
self.state.agent_pos = Coord(*self.board["_poc_home"])
ghost_home = Coord(*self.board["_ghost_home"])
for g in range(self.board["_num_ghosts"]):
pos = Coord(ghost_home.x + g % 2, ghost_home.y + g // 2)
self.state.ghosts.append(Ghost(pos, direction=-1))
self.state.food_pos = np.random.binomial(1,
config["_food_prob"],
size=self.grid.n_tiles + 1)
self.state.power_step = 0
return self.state
def __init__(self, board_size=7, num_rocks=8, use_heuristic=False):
assert board_size in list(
config.keys()) and num_rocks in config[board_size]['size']
self.num_rocks = num_rocks
self._use_heuristic = use_heuristic
self._rock_pos = [
Coord(*rock) for rock in config[board_size]['rock_pos']
]
self._agent_pos = Coord(*config[board_size]['init_pos'])
self.grid = Grid(board_size, board_size)
for idx, rock in enumerate(self._rock_pos):
self.grid.board[rock] = idx
self.action_space = Discrete(len(Action) + self.num_rocks)
self.observation_space = Discrete(len(Obs))
self._discount = .95
self._reward_range = 20
self._penalization = -100
self._query = 0
def _decode_state(self, state, as_array=False):
agent_pos = Coord(*state['agent_pos'])
rock_state = RockState(agent_pos)
for r in state['rocks']:
rock = Rock(pos=0)
rock.__dict__.update(r)
rock_state.rocks.append(rock)
if as_array:
rocks = []
for rock in rock_state.rocks:
rocks.append(rock.status)
return np.concatenate([[self.grid.get_index(agent_pos)], rocks])
return rock_state
def __dict2np__(self, state):
idx = self.grid.get_index(Coord(*state['agent_pos']))
rocks = []
for rock in state['rocks']:
rocks.append(rock['status'])
return np.concatenate([[idx], rocks])
def get_tag_coord(self, idx):
assert 0 <= idx < self.n_tiles
if idx < 20:
return Coord(idx % 10, idx // 10)
idx -= 20
return Coord(idx % 3 + 5, idx // 3 + 2)
def get_coord(self, idx):
assert idx >= 0 and idx < self.n_tiles
return Coord(idx % self.x_size, idx // self.x_size)
def __init__(self,
board_size=7,
num_rocks=8,
use_heuristic=False,
observation='o',
stay_inside=False):
"""
:param board_size: int board is a square of board_size x board_size
:param num_rocks: int number of rocks on board
:param use_heuristic: bool usage unclear
:param observation: str must be one of
'o': observed value only
'po': position of the agent + the above
'poa': the above + the action taken
"""
assert board_size in list(config.keys()) and \
num_rocks == len(config[board_size]["rock_pos"])
self.num_rocks = num_rocks
self._use_heuristic = use_heuristic
self._rock_pos = \
[Coord(*rock) for rock in config[board_size]['rock_pos']]
self._agent_pos = Coord(*config[board_size]['init_pos'])
self.grid = Grid(board_size, board_size)
for idx, rock in enumerate(self._rock_pos):
self.grid.board[rock] = idx
self.action_space = Discrete(len(Action) + self.num_rocks)
self._discount = .95
self._reward_range = 20
self._penalization = -100
self._query = 0
if stay_inside:
self._out_of_bounds_penalty = 0
else:
self._out_of_bounds_penalty = self._penalization
self.state = None
self.last_action = None
self.done = False
self.gui = None
assert observation in ['o', 'oa', 'po', 'poa']
if observation == 'o':
self._make_obs = lambda obs, a: obs
self.observation_space = Discrete(len(Obs))
elif observation == 'oa':
self._make_obs = self._oa
self.observation_space =\
Box(low=0,
high=np.append(max(Obs), np.ones(self.action_space.n)),
dtype=np.int)
elif observation == 'po':
self._make_obs = self._po
self.observation_space = \
Box(low=0,
high=np.append(np.ones(self.grid.n_tiles), max(Obs)),
dtype=np.int)
elif observation == 'poa':
self._make_obs = self._poa
self.observation_space = \
Box(low=0,
high=np.concatenate((np.ones(self.grid.n_tiles),
[max(Obs)],
np.ones(self.action_space.n))),
dtype=np.int)
