coach运行流程梳理

env_instance = create_environment(tuning_parameters)
agent = eval(tuning_parameters.agent.type + '(env_instance, tuning_parameters)')

def create_environment(tuning_parameters):
    env_type_name, env_type = EnvTypes().verify(tuning_parameters.env.type)
    env = eval(env_type)(tuning_parameters)
    return env

class GymEnvironmentWrapper(EnvironmentWrapper):
    def __init__(self, tuning_parameters):
        EnvironmentWrapper.__init__(self, tuning_parameters)

        # env parameters

class EnvironmentWrapper(object):
    def __init__(self, tuning_parameters):
        """
        :param tuning_parameters:
        :type tuning_parameters: Preset
        """
        # env initialization
        self.game = []
        self.actions = {}
        self.state = []
        self.reward = 0
        self.done = False
        self.default_action = 0
        self.last_action_idx = 0
        self.episode_idx = 0
        self.last_episode_time = time.time()
        self.info = []
        self.action_space_low = 0
        self.action_space_high = 0
        self.action_space_abs_range = 0
        self.actions_description = {}
        self.discrete_controls = True
        self.action_space_size = 0
        self.key_to_action = {}
        self.width = 1
        self.height = 1
        self.is_state_type_image = True
        self.measurements_size = 0
        self.phase = RunPhase.TRAIN
        self.tp = tuning_parameters
        self.record_video_every = self.tp.visualization.record_video_every
        self.env_id = self.tp.env.level
        self.video_path = self.tp.visualization.video_path
        self.is_rendered = self.tp.visualization.render
        self.seed = self.tp.seed
        self.frame_skip = self.tp.env.frame_skip
        self.human_control = self.tp.env.human_control
        self.wait_for_explicit_human_action = False
        self.is_rendered = self.is_rendered or self.human_control
        self.game_is_open = True
        self.renderer = Renderer()

    @property
    def measurements(self):
        assert False

    @measurements.setter
    def measurements(self, value):
        assert False

    @property
    def observation(self):
        assert False

    @observation.setter
    def observation(self, value):
        assert False

    def _idx_to_action(self, action_idx):
        """

class EnvRegistry(object):
    """Register an env by ID. IDs remain stable over time and are

if 'action_intrinsic_reward' in action_info.keys():

class PolicyOptimizationAgent(Agent):

agent = eval(tuning_parameters.agent.type + '(env_instance, tuning_parameters)')
# Start the training or evaluation
if tuning_parameters.evaluate:
    agent.evaluate(sys.maxsize, keep_networks_synced=True)  # evaluate forever
else:
    agent.improve()