def __init__(self, observation_settings, opponent=None,
reset_arm_after_move=True):
"""Initializes a `TicTacToe` task.
Args:
observation_settings: An `observations.ObservationSettings` namedtuple
specifying configuration options for each category of observation.
opponent: TicTacToeOpponent used for generating opponent moves.
reset_arm_after_move: Whether to reset arm to random position after every
piece being placed on the board.
"""
game_logic = tic_tac_toe_logic.TicTacToeGameLogic()
if opponent is None:
opponent = tic_tac_toe_logic.TicTacToeRandomOpponent()
markers = pieces.Markers(num_per_player=5,
observable_options=observations.make_options(
observation_settings,
observations.MARKER_OBSERVABLES))
self._reset_arm_after_move = reset_arm_after_move
super(TicTacToe, self).__init__(observation_settings=observation_settings,
opponent=opponent,
game_logic=game_logic,
board=boards.CheckerBoard(),
markers=markers)
def test_invalid_player_id(self):
markers = pieces.Markers(num_per_player=5)
physics = mjcf.Physics.from_mjcf_model(markers.mjcf_model)
invalid_player_id = 99
with self.assertRaisesWithLiteralMatch(
ValueError, pieces._INVALID_PLAYER_ID.format(1, 99)):
markers.mark(physics=physics, player_id=invalid_player_id, pos=(1, 2, 3))
def test_too_many_moves(self):
num_per_player = 5
player_id = 0
markers = pieces.Markers(num_per_player=num_per_player)
physics = mjcf.Physics.from_mjcf_model(markers.mjcf_model)
for _ in range(num_per_player):
markers.mark(physics=physics, player_id=player_id, pos=(1, 2, 3))
with self.assertRaisesWithLiteralMatch(
RuntimeError,
pieces._NO_MORE_MARKERS_AVAILABLE.format(num_per_player, player_id)):
markers.mark(physics=physics, player_id=player_id, pos=(1, 2, 3))
def test_position_observable(self):
num_per_player = 3
markers = pieces.Markers(num_per_player=num_per_player)
physics = mjcf.Physics.from_mjcf_model(markers.mjcf_model)
all_positions = [
[(0, 1, 2), (3, 4, 5), (6, 7, 8)], # Player 0
[(-1, 2, -3), (4, -5, 6)], # Player 1
]
for player_id, positions in enumerate(all_positions):
for marker_pos in positions:
markers.mark(physics=physics, player_id=player_id, pos=marker_pos)
expected_positions = np.zeros((2, num_per_player, 3), dtype=np.double)
expected_positions[0, :len(all_positions[0])] = all_positions[0]
expected_positions[1, :len(all_positions[1])] = all_positions[1]
observed_positions = markers.observables.position(physics)
np.testing.assert_array_equal(
expected_positions.reshape(-1, 3), observed_positions)
def __init__(self, board_size, observation_settings, opponent=None,
reset_arm_after_move=True):
"""Initializes a `Go` task.
Args:
board_size: board size
observation_settings: An `observations.ObservationSettings` namedtuple
specifying configuration options for each category of observation.
opponent: Go opponent to use for the opponent player actions.
reset_arm_after_move: Whether to reset arm to random position after every
piece being placed on the board.
"""
game_logic = go_logic.GoGameLogic(board_size=board_size)
if opponent is None:
opponent = go_logic.GoGTPOpponent(board_size=board_size,
mixture_p=_DEFAULT_OPPONENT_MIXTURE)
self._last_valid_move_is_pass = False
super(Go, self).__init__(observation_settings=observation_settings,
opponent=opponent,
game_logic=game_logic,
board=boards.GoBoard(boardsize=board_size),
markers=pieces.Markers(
player_colors=(_BLACK, _WHITE),
halfwidth=_GO_PIECE_SIZE,
num_per_player=board_size*board_size*2,
observable_options=observations.make_options(
observation_settings,
observations.MARKER_OBSERVABLES),
board_size=board_size))
self._reset_arm_after_move = reset_arm_after_move
# Add an observable exposing the move history (to reconstruct game states)
move_history_observable = observable.Generic(
lambda physics: self._game_logic.get_move_history())
move_history_observable.configure(
**observation_settings.board_state._asdict())
self._task_observables['move_history'] = move_history_observable