8896f02f94da16afe5b3ea5330e4f99245b75d6b,examples/lqr_pg.py,,experiment,#Any#Any#Any#Any#,24

Before Change

        dataset_eval = core.evaluate(n_episodes=ep_per_run)
        print("policy parameters: ", policy.get_weights())
        J = compute_J(dataset_eval, gamma=mdp.info.gamma)
        print("J at iteration " + str(i) + ": " + str(np.mean(J)))


if __name__ == "__main__":

After Change

def experiment(alg, n_epochs, n_iterations, ep_per_run):
    np.random.seed()

    logger = Logger(alg.__name__, results_dir=None)
    logger.strong_line()
    logger.info("Experiment Algorithm: " + alg.__name__)

    // MDP
    mdp = LQR.generate(dimensions=1)

    approximator = Regressor(LinearApproximator,
                             input_shape=mdp.info.observation_space.shape,
                             output_shape=mdp.info.action_space.shape)

    sigma = Regressor(LinearApproximator,
                      input_shape=mdp.info.observation_space.shape,
                      output_shape=mdp.info.action_space.shape)

    sigma_weights = 2 * np.ones(sigma.weights_size)
    sigma.set_weights(sigma_weights)

    policy = StateStdGaussianPolicy(approximator, sigma)

    // Agent
    optimizer = AdaptiveOptimizer(eps=.01)
    algorithm_params = dict(optimizer=optimizer)
    agent = alg(mdp.info, policy, **algorithm_params)

    // Train
    core = Core(agent, mdp)
    dataset_eval = core.evaluate(n_episodes=ep_per_run)
    J = compute_J(dataset_eval, gamma=mdp.info.gamma)
    logger.epoch_info(0, J=np.mean(J), policy_weights=policy.get_weights())

    for i in trange(n_epochs, leave=False):
        core.learn(n_episodes=n_iterations * ep_per_run,
                   n_episodes_per_fit=ep_per_run)
        dataset_eval = core.evaluate(n_episodes=ep_per_run)
        J = compute_J(dataset_eval, gamma=mdp.info.gamma)
        logger.epoch_info(i+1, J=np.mean(J), policy_weights=policy.get_weights())


if __name__ == "__main__":
    algs = [REINFORCE, GPOMDP, eNAC]

Instances