Too slow search using MCTS in OpenAI Atari games

Question

I'm recently using Monte Carlo Tree Search in OpenAi Gym Atari, but the result isn't satisfying.

Without render, the game lasts about 180 steps ( env.step() was called this much time ) with random agent. However, my MCTS agent only made the game last 12 steps. And it took pretty much time to give a next step.

I guess it's the problem of rollout. I build the MCTS tree using nodes containing AtariEnv objects, and deepcopy it each time I rollout, add the reward. So it takes about 1 second to expand nodes and rollout, if I do 100 iterations, that would be massive waiting time.

My code of rollout is shown below:

def rollout_(current,if_render):
        '''
        current is going to be a Node object
        '''
        sandBox = deepcopy(current.state)
        endReward = 0
        done = False
        while done != True:
            action=sandBox.action_space.sample()
            _,reward,done,info = sandBox.step(action)# wierd return obs_next
            if reward > 0:
                reward *= 2
            endReward += reward-0.008
        return endReward

Anyone can help?