policygradient.py 文件源码

def _initialize(self):
        logger.debug("Initializing Policy.")
        # check if policy is already initialized by the user
        if self.policy.initialized:
            logger.debug("Use pre-set policy parameters.")
            return self.policy.parameters

        # outerwise draw an element at random from the parameter space
        parameter = self.parameter_space.sample()

        for _ in range(1000):
            self.policy.parameters = parameter
            grad = self.estimator(self.policy)

            if (norm(grad) >= 1000 * self.eps):
                return parameter

            parameter = self.parameter_space.sample()

        logger.error('Unable to find non-zero gradient.')