Method _episode

numpy_ml/rl_models/agents.py:992–1052

Run or train the agent on an episode. Parameters ---------- max_steps : int The maximum number of steps to run the episode. render : bool Whether to render the episode during training. update : bool Whether to perf

(self, max_steps, render, update=True)

        return total_rwd, n_steps

    def _episode(self, max_steps, render, update=True):
        """
        Run or train the agent on an episode.

        Parameters
        ----------
        max_steps : int
            The maximum number of steps to run the episode.
        render : bool
            Whether to render the episode during training.
        update : bool
            Whether to perform the Q function backups after each step. Default
            is True.

        Returns
        -------
        reward : float
            The total reward on the episode.
        steps : float
            The number of steps taken on the episode.
        """
        self.flush_history()

        obs = self.env.reset()
        HS = self.episode_history

        action = self.act(obs)
        s = self._obs2num[obs]
        a = self._action2num[action]

        # store initial (state, action) tuple
        HS["state_actions"].append((s, a))

        total_reward, n_steps = 0.0, 0
        for i in range(max_steps):
            if render:
                self.env.render()

            # take action
            obs, reward, done, info = self.env.step(action)
            n_steps += 1

            # record rewards
            HS["rewards"].append(reward)
            total_reward += reward

            # generate next state and action
            action = self.act(obs)
            s_ = self._obs2num[obs] if not done else None
            a_ = self._action2num[action]

            # store next (state, action) tuple
            HS["state_actions"].append((s_, a_))

            if update:
                self.update()

            if done: