test_discrete.py 文件源码

def test_sarsa():
    env = TwoRoundDeterministicRewardEnv()
    np.random.seed(123)
    env.seed(123)
    random.seed(123)
    nb_actions = env.action_space.n

    # Next, we build a very simple model.
    model = Sequential()
    model.add(Dense(16, input_shape=(1,)))
    model.add(Activation('relu'))
    model.add(Dense(nb_actions, activation='linear'))

    policy = EpsGreedyQPolicy(eps=.1)
    sarsa = SARSAAgent(model=model, nb_actions=nb_actions, nb_steps_warmup=50, policy=policy)
    sarsa.compile(Adam(lr=1e-3))

    sarsa.fit(env, nb_steps=20000, visualize=False, verbose=0)
    policy.eps = 0.
    h = sarsa.test(env, nb_episodes=20, visualize=False)
    assert_allclose(np.mean(h.history['episode_reward']), 3.)