def min_max_alpha_beta(game, turn, max_depth, alpha=-sys.float_info.max, beta=sys.float_info.max):
best_score_move = None
game.available_moves()
moves = game.moves
if not moves:
return 0, None
for move in moves:
new_game = Checkers()
new_game.board_state = game.board_state
new_game.turn = game.turn
new_game.moves_queen_with_out_capture = game.moves_queen_with_out_capture
new_game.move(move)
winner = new_game.win
if winner != 0:
return winner*10000, move
else:
if max_depth <= 1:
score = evaluate(new_game)
else:
score, _ = min_max_alpha_beta(new_game, -turn, max_depth-1, alpha, beta)
if turn > 0:
if score > alpha:
alpha = score
best_score_move = move
else:
if score < beta:
beta = score
best_score_move = move
if alpha >= beta:
break
return alpha if turn > 0 else beta, best_score_move
from dqn import DQN from coach import Coach from game import Checkers from arena import Arena import torch from tqdm import tqdm game = Checkers() dqn = DQN() NUM_EPOCHS = 80 NUM_TRAIN_GAMES = 200 NUM_EVAL_GAMES = 20 HALF = int(NUM_EVAL_GAMES/2) TRAIN_EPSILON = 0.4 GAMMA = 0.8 EVAL_EPSILON = 0.01 LR = 0.0001 PATH = "models/test.pt" coach = Coach( game=game, nnet=dqn, lr=LR) num_improvements = 0 for _ in range(NUM_EPOCHS): ## Saving a copy of the current network weights to pnet torch.save(coach.nnet.state_dict(), PATH) coach.pnet.load_state_dict(torch.load(PATH))
from game import Checkers
from dqn import DQN
import torch
PATH = "models/30Improve.pt"
dqn = DQN()
dqn.load_state_dict(torch.load(PATH))
game = Checkers()
game.reset()
board = game.board
## Display
print(board)
color = 'red'
while not game.get_game_ended(board, color):
actions = game.get_possible_actions(board, color)
state = game.get_state(board, color)
state_actions = game.make_inputs(state, actions)
if color == 'black':
print("Agent turn")
print("Agent possible actions")
out = dqn(state_actions)
for a, q in zip(actions, out):
print(game.flip_action(a), q.item())
print(torch.max(out))
idx = torch.argmax(out)
action = actions[idx]
with tf.Session() as session:
model1.set_session(session)
model2.set_session(session)
session.run(tf.global_variables_initializer())
GAME.reset()
wins = 0
for i in range(MIN_EXPERIENCES):
GAME.available_moves()
if GAME.win != 0:
GAME.reset()
move = random_play(GAME)
action = encoding_move(move)
GAME.move(move)
if GAME.win == 0:
new_GAME = Checkers()
new_GAME.board_state = np.array(GAME.board_state)
new_GAME.turn = GAME.turn
new_GAME.moves_queen_with_out_capture = GAME.moves_queen_with_out_capture
move = min_max.min_max_player(new_GAME, new_GAME.turn)
GAME.move(move)
reward = GAME.win
experience_replay_buffer.add_experince(action, GAME.board_state,
reward)
t0 = datetime.now()
for i in range(num_episodes):
total_t, episode_reward, duration, num_steps_in_episode, time_per_step, epsilon = play_one(
total_t, experience_replay_buffer, model1, model2, epsilon)
episode_rewards[i] = episode_reward
def setUp(self):
self.test_Game = Checkers()
class TestBoard(unittest.TestCase):
def setUp(self):
self.test_Game = Checkers()
def test_move(self):
self.test_Game.move((5, 0, 4, 1))
print('move')
expected_result = np.array([[0, -1, 0, -1, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, -1, 0],
[0, -1, 0, -1, 0, -1, 0, -1],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 1, 0]])
np.testing.assert_array_equal(expected_result,
self.test_Game.board_state)
self.test_Game.reset()
print("capture")
self.test_Game.board_state = np.array([[0, -1, 0, -1, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, -1, 0],
[0, -1, 0, 0, 0, -1, 0, -1],
[0, 0, -1, 0, -1, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 1, 0]])
self.test_Game.move((4, 1, 2, 3))
expected_result = np.array([[0, -1, 0, -1, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, -1, 0],
[0, -1, 0, 1, 0, -1, 0, -1],
[0, 0, 0, 0, -1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 1, 0]])
np.testing.assert_array_equal(expected_result,
self.test_Game.board_state)
self.test_Game.reset()
print("trying no capture")
self.test_Game.board_state = np.array([[0, -1, 0, -1, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, -1, 0],
[0, -1, 0, 0, 0, -1, 0, -1],
[0, 0, -1, 0, -1, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 1, 0]])
self.test_Game.move((4, 1, 0, 3))
expected_result = np.array([[0, -1, 0, -1, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, -1, 0],
[0, -1, 0, 0, 0, -1, 0, -1],
[0, 0, -1, 0, -1, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 1, 0]])
np.testing.assert_array_equal(expected_result,
self.test_Game.board_state)
self.test_Game.reset()
print("capture from first col")
self.test_Game.turn = -1
self.test_Game.board_state = np.array([[0, -1, 0, -1, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, -1, 0],
[0, -1, 0, 0, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 1, 0, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 1, 0]])
self.test_Game.move((3, 0, 5, 2))
expected_result = np.array([[0, -1, 0, -1, 0, -1, 0, -1],
[-1, 0, -1, 0, -1, 0, -1, 0],
[0, -1, 0, 0, 0, -1, 0, -1],
[0, 0, -1, 0, -1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, -1, 0, 1, 0, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 1, 0]])
np.testing.assert_array_equal(expected_result,
self.test_Game.board_state)