Made some test cases less flaky

    learner = SineLearner()

    optimizer = dc.models.tensorgraph.optimizers.Adam(learning_rate=5e-3)

    maml = dc.metalearning.MAML(learner, meta_batch_size=4, optimizer=optimizer)

    maml.fit(5000)

    maml.fit(7000)

    # Test it out on some new tasks and see how it works.

        TestPolicy(),

        max_rollout_length=20,

        optimizer=Adam(learning_rate=0.003))

    ppo.fit(50000)

    ppo.fit(80000)

    # It should have learned that the expected value is very close to zero, and that the best

    # action is to walk away.

    action_prob, value = ppo.predict([[0]])

    assert -0.5 < value[0] < 0.5

    assert -0.8 < value[0] < 0.5

    assert action_prob.argmax() == 37

    assert ppo.select_action([[0]], deterministic=True) == 37