def observation(self, board):
"""Convert observation to numpy array with a unique channel for each tile.
A `Board` cannot be used as an observaton. RLlib will complain and crash because
RLlib expects arrays as observations. Therefore, we convert the `Board` to a
numpy array, where the first channel has value 1 if it's empty. The second
channel correspond to tiles with value 2, the third with value 3 and so on.
The number of channels in the observation will be `1 + log2(max_tile_value)`.
For example, `max_tile_value == 256` --> we have 9 tile values.
Note:
We assume all tiles are a multiple of 2!
Returns:
A dict with the following keys and values:
- 'valid_action_mask': np.ndarray(4, float)
The available actions
- 'board': np.ndarray((n_rows, n_cols, n_channels), float)
The board (in one-hot format).
"""
channel_indices = np.log2(np.where(board.values == 0, 1, board.values))
frac_values, _ = np.modf(channel_indices)
if not frac_values.max() == 0:
raise ValueError(
"Unexpected input: got a tile that was not a power of 2. Can't "
"safely convert observation.")
channel_indices = channel_indices.astype(int)
yy, xx = np.meshgrid(*[range(dim) for dim in channel_indices.shape])
one_hot_board = np.zeros(self.env.observation_space["board"].shape)
if K.image_data_format() == "channels_first":
one_hot_board[channel_indices.ravel(),
yy.ravel(),
xx.ravel()] = 1.0
else:
one_hot_board[yy.ravel(),
xx.ravel(),
channel_indices.ravel()] = 1.0
valid_action_mask = np.zeros(4)
for action in Board.get_available_actions(board):
index = action.value - 1 # enums are 1-indexed, so we subtract by 1.
valid_action_mask[index] = 1.0
processed_obs = {
"valid_action_mask": valid_action_mask,
"board": one_hot_board
}
return processed_obs
def test_available_actions_work_as_expected():
s = """1 2 3
1 4 5
6 7 8"""
board = board_from_string(s)
available_actions = Board.get_available_actions(board)
assert available_actions == set([Action.DOWN, Action.UP])
s = """ 1 2 3 4
5 6 7 8
9 10 11 12
13 14 15 16"""
board = board_from_string(s)
available_actions = Board.get_available_actions(board)
assert available_actions == set([])
s = """ 1 2 2 4
5 2 7 8
9 10 11 12
13 14 15 16"""
board = board_from_string(s)
available_actions = Board.get_available_actions(board)
assert available_actions == set([Action.LEFT, Action.RIGHT, Action.UP, Action.DOWN])