class GameEnv(object):
def __init__(self, level='env/level.csv'):
self.game = Game(level)
self.repeat_frame_skip = 4
def reset(self):
self.game.reset()
state = self.game.state()
self.agent_coord = state['coord']
return state
def step(self, action):
for _ in range(self.repeat_frame_skip):
self.game.step(action)
state = self.game.state()
dead = state['dead']
goal = state['goal']
coord = state['coord']
reward = -1 + (coord[0] -
self.agent_coord[0]) + 100 * goal - 100 * dead
done = dead or goal
self.agent_coord = coord
return state, reward, done, {
'goal': goal,
'dead': dead,
'distance': self.agent_coord[0]
}
def render(self, mode='rgb_array'):
pixels = self.game.render(mode)
pixels = np.swapaxes(pixels, 0, 1)
return pixels
if __name__ == "__main__":
start_iter = 50000
init_checkpoint = None
num_epochs = 2000001
dim_states = 52
rl = RL(dim_states, lr_a=0.0001, lr_c=0.0001, init_checkpoint=init_checkpoint)
# fine-tune
if start_iter != 0 and not init_checkpoint:
rl.load_model('rl', start_iter)
env = Game()
for episode in range(start_iter, num_epochs):
env.reset()
print()
history_vec = []
history_pid = []
while 1:
pid = env.now_player_id
# 无人叫地主 or 游戏结束，记录所有存档
if env.landlord_count == 3 or env.winner >= 0:
for i in range(3):
state, f_reward, y_reward, act_ids, dyn_vec, _, label_mask, attn_mask = env.observe(pid)
print('玩家', pid, '获得奖励', y_reward)
pid = (pid + 1) % 3
env.now_player_id = pid
break