def add_game(self, player_strategy, rnd, starting_game_position=None):
"""Runs a game with the given strategy and randomness source, then
enrolls the outcome in the dataset.
If @p starting_position is a Game object, start from that position.
Returns the number of examples (moves) added.
"""
states = np.zeros((1, EXAMPLE_WIDTH))
num_moves = 0
game = starting_game_position or Game(rnd=rnd)
running = True
while running:
intermediate_board, turn_outcome = (
game.do_turn_and_retrieve_intermediate(
player_strategy.get_move(game.board(), game.score())))
running = (turn_outcome != GAMEOVER)
num_moves += (turn_outcome != ILLEGAL)
if turn_outcome == OK:
states = np.append(states,
Board.as_vector(intermediate_board),
axis=0)
self._num_examples += 1
player_strategy.notify_outcome(game.board(), game.score())
scores = Dataset.evaluate_states(states, game.board(), game.score)
assert(len(states) == len(scores))
batch_size_so_far = self._example_batches[-1].shape[0]
if len(states) + batch_size_so_far > MAX_BATCH_SIZE:
self._example_batches.append(np.zeros((0, EXAMPLE_WIDTH)))
self._score_batches.append(np.zeros((0, 1)))
self._example_batches[-1] = \
np.append(self._example_batches[-1], states, axis=0)
self._score_batches[-1] = np.append(self._score_batches[-1], scores)
return len(states)
