def main(self):
end, winner, _ = False, None, -1
game = Conn4Game(self.rows, self.cols)
input_dim = game.encoded_board_dim()
output_dim = game.action_size()
net = Conn4Net(input_dim, output_dim, load=True, fname=self.net_file)
mcts = MCTS(net)
NSIMS = 500
TEMP = 0
while not end:
if self.human_side == game.currentPlayer:
print(game)
print('Enter your move (column):')
move = int(input())
game.apply_action(move)
else:
p = mcts.get_action_prob(game, NSIMS, temp=TEMP)
move = np.argmax(p)
game.apply_action(move)
print('AI just played on column {}'.format(move))
end, winner, _ = game.game_ended(1)
print(game)
if winner == 1 and self.human_side == 1:
print('You win!')
elif winner == 1 and self.human_side == 2:
print('You lose!')
elif winner == 2 and self.human_side == 2:
print('You win!')
elif winner == 2 and self.human_side == 1:
print('You lose!')
else:
print('Draw!')
def run_match(net1, net2, n_sims, n_games, rows=6, cols=7):
print('Running match between {} and {}'.format(net1, net2))
game = Conn4Game(rows=rows, cols=cols)
input_dim = game.encoded_board_dim()
output_dim = game.action_size()
_net1 = Conn4Net(input_dim, output_dim, load=True, fname=net1)
_net2 = Conn4Net(input_dim, output_dim, load=True, fname=net2)
pit = Conn4Pit(_net1, _net2, n_sims, n_games, rows, cols)
score = pit.run()
return net1, net2, score
def play_net2_net1(self, ngames, mcts1, mcts2):
wins_net1 = 0
draws = 0
for _ in range(ngames):
game = Conn4Game(self.rows, self.cols)
end, winner, _ = False, None, -1
while not end:
player = self.net1 if game.currentPlayer == 2 else self.net2
mcts = mcts1 if player == self.net1 else mcts2
p = mcts.get_action_prob(game, self.n_sims, temp=0)
a = np.argmax(p)
game.apply_action(a)
end, winner, _ = game.game_ended(1)
if winner == 2:
wins_net1 += 1
elif winner is None:
draws += 1
return wins_net1 + draws * 0.5
def execute_episode(self): examples = [] game = Conn4Game(self.rows, self.cols) dim = len(game.board) mcts = MCTS(self.best_nn) move = 0 while True: temp = 1#1 if move < 10 else 0.1 p_a = mcts.get_action_prob(game, self.n_sims, temp=temp) s_h = game.hash() s_x = game.encode_board() # filter nonsense moves #p_a = [p_a[i] if game.board[int(i/dim)][i%dim] == -1 else 0 for i in range(len(p_a))] #p_a = [pi/sum(p_a) for pi in p_a] # filter nonsense moves examples.append([s_x, p_a, None]) a = np.random.choice(range(len(p_a)), p=p_a) game.apply_action(a) end, winner, r = game.game_ended(game.currentPlayer) if end: # does the first player won? if winner == 1: m = -1 elif winner == 2: m = 1 else: m = 0 for ex in examples: ex[-1] = m * r m*= -1 return examples move+= 1
if end: # does the first player won? if winner == 1: m = -1 elif winner == 2: m = 1 else: m = 0 for ex in examples: ex[-1] = m * r m*= -1 return examples move+= 1 if __name__ == '__main__': rows = 6 cols = 7 fname = None#'models_conv/best_model6x7_93.hdf5' game = Conn4Game(rows, cols) input_dim = game.encoded_board_dim() output_dim = game.action_size() net1 = Conn4Net(input_dim, output_dim, load=fname is not None, fname=fname) p = PolicyIter(nnet=net1, n_iters=5000, n_episodes=10, n_sims=40, n_games_pit=4, batch_size=32, win_eps=0.55, rows=rows, cols=cols) p.policy_iter()