def test_game_over_checking_works_as_expected():
not_game_over_strings = ["1 1 1 1", "0 0 0 0", "0 1 0 0", "3 2 1 3 4 5 6 7 8"]
for s in not_game_over_strings:
board = board_from_string(s)
assert not Board.is_game_over(board)
game_over_strings = ["1 2 3 4", "1 2 3 4 5 6 7 8 9", "4 3 7 2"]
for s in game_over_strings:
board = board_from_string(s)
assert Board.is_game_over(board)
def step(self, action: int):
"""A single step in the game
rewards: the natural logarithm of difference of the 2048 scoring.
We also add some penalty in case of no-op
"""
self._step_counter += 1
done = False
if self._step_counter > 100 and self._step_counter / (self.t + 1) > 5:
# add 1 to avoid DivisionByZero for `self.t`. Yes, it happenend.
done = True
action = Action(1 + action) # discrete -> enum (which is 1-indexed)
modified_board = Board.apply_action_on_board(self.board, action)
info = {"step": self.t}
if Board.is_game_over(modified_board):
done = True
reward = 0
else:
# An action is invalid if it doesn't change the board.
valid_action = modified_board != self.board
if not valid_action:
# We penalize the agent for doing no-op moves!!! >:(
penalty = -0.1
info["no-op"] = True
else:
modified_board = Board.spawn_random_tile(modified_board)
penalty = 0
self.t += 1
info["no-op"] = False
diff = modified_board.score - self.board.score
reward = np.log(1 + diff) + penalty
reward = np.clip(reward, -11, 10) # TODO: move to a wrapper.
self.board = modified_board
return self.board, reward, done, info