DQN

dqn.py

class DQN(DQNBase):
    def __init__(self):
        self.k = HP['frame_skipping']

    def initialState(self):
        states = []
        obs = self.env.reset()
        obs = self.model.preprocess(obs)
        for _ in range(HP['stacked_frame_size']):
            states.append(obs)
        return np.dstack(tuple(states))

    def executeAction(self, action, state):
        if self.k == HP['frame_skipping']:
            # new action
            self.k = 0
            self.last_action = action
        else:
            # repeat last action
            self.k += 1
            action = self.last_action
        s1, reward, done, _ = self.env.step(action)
        newObservation = self.model.preprocess(s1)
        newState = state[:, :, 1:]
        newState = np.dstack((newState, newObservation))
        return {'state': state,
               'action': action,
               'reward': reward,
               'next_state': newState,
               'done': done }