def update_env_after_move(board: Board, move: Move, north_moved):
if not KalahEnvironment.is_permitted(board, move, north_moved):
raise Exception('Move not permitted')
if move.index == 0:
KalahEnvironment.swap_sides(board)
return Side.opposite(move.side)
seeds_to_sow = board.get_seeds(move.side, move.index)
board.set_seeds(move.side, move.index, 0)
holes = board.holes
receiving_holes = 2 * holes + 1
rounds = seeds_to_sow // receiving_holes
remaining_seeds = seeds_to_sow % receiving_holes
if rounds != 0:
for hole in range(1, holes + 1):
board.add_seeds(Side.NORTH, hole, rounds)
board.add_seeds(Side.SOUTH, hole, rounds)
board.add_seeds_to_store(move.side, rounds)
sow_side = move.side
sow_hole = move.index
for _ in range(remaining_seeds):
sow_hole += 1
if sow_hole == 1:
sow_side = Side.opposite(sow_side)
if sow_hole > holes:
if sow_side == move.side:
sow_hole = 0
board.add_seeds_to_store(sow_side, 1)
continue
else:
sow_side = Side.opposite(sow_side)
sow_hole = 1
board.add_seeds(sow_side, sow_hole, 1)
if sow_side == move.side and sow_hole > 0 and board.get_seeds(sow_side, sow_hole) == 1 \
and board.get_seeds_op(sow_side, sow_hole) > 0:
board.add_seeds_to_store(move.side, 1 + board.get_seeds_op(sow_side, sow_hole))
board.set_seeds(move.side, sow_hole, 0)
board.set_seeds_op(move.side, sow_hole, 0)
game_over = KalahEnvironment.game_finished(board)
if game_over:
finished_side = Side.NORTH if KalahEnvironment.side_has_no_seeds(board, Side.NORTH) else Side.SOUTH
seeds = 0
collecting_side = Side.opposite(finished_side)
for hole in range(1, board.holes + 1):
seeds += board.get_seeds(collecting_side, hole)
board.set_seeds(collecting_side, hole, 0)
board.add_seeds_to_store(collecting_side, seeds)
if sow_hole == 0 and (move.side == Side.NORTH or north_moved):
return move.side
return Side.opposite(move.side)
def h1(state: KalahEnvironment, side: Side) -> float:
my_mancala = state.board.get_seeds_in_store(side)
opponent_mancala = state.board.get_seeds_in_store(Side.opposite(side))
diff = my_mancala - opponent_mancala
return diff
def set_seeds_op(self, side: Side, hole: int, seeds: int):
if hole < 1 or hole > self.holes:
raise ValueError(
'Hole number must be between 1 and number of holes')
if seeds < 0:
raise ValueError('There has to be a non-negative number of seeds')
self.board[Side.get_index(Side.opposite(side))][self.holes + 1 -
hole] = seeds
def get_winner(self) -> Side or None:
if not self.has_game_ended():
raise Exception('Game has not ended')
last_move_side = Side.NORTH if KalahEnvironment.side_has_no_seeds(self.board, Side.NORTH) else Side.SOUTH
other_side = Side.opposite(last_move_side)
last_move_side_seeds = self.board.get_seeds_in_store(other_side)
for hole in range(1, self.board.holes + 1):
last_move_side_seeds += self.board.get_seeds(other_side, hole)
other_side_seeds = self.board.get_seeds_in_store(last_move_side)
if other_side_seeds > last_move_side_seeds:
return last_move_side
elif other_side_seeds < last_move_side_seeds:
return other_side
else:
return None
def env_runner(env, trainer_side, ac_net, opp_agent):
"""
The logic of the thread runner. In brief, it constantly keeps on running
the policy, and as long as the rollout exceeds a certain length, the thread
runner appends the policy to the queue.
"""
rollout = Rollout()
while not env.has_game_ended():
# There is no choice if only one action is left. Taking that action automatically must be seen as
# a characteristic behaviour of the environment. This helped the learning of the agent
# to be more numerically stable (this is an empirical observation).
if len(env.get_valid_moves()) == 1:
action_left_to_perform = env.get_valid_moves()[0]
env.do_move(action_left_to_perform)
continue
if env.side_to_play == trainer_side:
# If the agent is playing as NORTH, it's input would be a flipped board
flip_board = env.side_to_play == Side.NORTH
state = env.board.get_board_image(flipped=flip_board)
mask = env.get_mask()
action, value = ac_net.sample(state, mask)
# Because the pie move with index 0 is ignored, the action indexes must be shifted by one
reward = env.do_move(Move(trainer_side, action + 1))
rollout.add(state, action, reward, value, mask)
else:
assert env.side_to_play == Side.opposite(trainer_side)
action = opp_agent.produce_action(env.board.get_board_image(),
env.get_mask(), env.side_to_play)
env.do_move(Move(env.side_to_play, action + 1))
# We replace the partial reward of the last move with the final reward of the game
final_reward = env.calculate_score_diff(trainer_side)
rollout.update_last_reward(final_reward)
if env.get_winner() == trainer_side:
rollout.add_win()
return rollout
def calculate_score_diff(self, side: Side):
diff = self.board.get_seeds_in_store(side) - self.board.get_seeds_in_store(Side.opposite(side))
return diff
def do_move(self, move: Move):
if move.index == 0:
self.my_side = Side.opposite(self.my_side)
self.side_to_play = KalahEnvironment.update_env_after_move(self.board, move, self.north_has_moved)
if move.side == Side.NORTH:
self.north_has_moved = True
def get_seeds_op(self, side: Side, hole: int):
if hole < 1 or hole > self.holes:
raise ValueError(
'Hole number must be between 1 and number of holes')
return self.board[Side.get_index(Side.opposite(side))][self.holes + 1 -
hole]