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
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
last_100_avg = episode_rewards[max(0, i - 100):i + 1].mean()
print("Episode:", i, "Duration:", duration, "Num steps:",
num_steps_in_episode, "Reward:", episode_reward,