Method greedy_policy

numpy_ml/rl_models/agents.py:1254–1301

Execute a deterministic greedy policy using the current agent parameters. Parameters ---------- max_steps : int The maximum number of steps to run the episode. render : bool Whether to render the episode during execution.

(self, max_steps, render=True)

        return self.behavior_policy(s)

    def greedy_policy(self, max_steps, render=True):
        """
        Execute a deterministic greedy policy using the current agent
        parameters.

        Parameters
        ----------
        max_steps : int
            The maximum number of steps to run the episode.
        render : bool
            Whether to render the episode during execution.

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

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

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

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

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

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

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

            if done:
                break

        return total_reward, n_steps


class DynaAgent(AgentBase):