def test_seed_not_set(self):
seed = 123
env1 = TestEnv()
env1.seed(seed)
brain1 = TestBrain(self._n_states, self._n_actions)
acting1 = EpsGreedyPolicy(self._eps)
replay_memory1 = Memory(self._capacity, self._batch_size)
agent1 = TdAgent(self._n_episodes, env1, brain1, acting1,
replay_memory1, self._gamma)
rewards1, _ = agent1.run()
env2 = TestEnv()
env2.seed(seed)
brain2 = TestBrain(self._n_states, self._n_actions)
acting2 = EpsGreedyPolicy(self._eps)
replay_memory2 = Memory(self._capacity, self._batch_size)
agent2 = TdAgent(self._n_episodes, env2, brain2, acting2,
replay_memory2, self._gamma)
rewards2, _ = agent2.run()
# assert rewards1 != rewards2
diff = rewards1 - rewards2
diff = np.abs(diff)
diff = np.sum(diff)
self.assertGreater(diff, 0)
def test_seed_set(self):
seed = 123
env1 = TestEnv()
env1.seed(seed)
brain1 = TestBrain(self._n_states, self._n_actions)
acting1 = EpsGreedyPolicy(self._eps)
replay_memory1 = Memory(self._capacity, self._batch_size)
agent1 = TdAgent(self._n_episodes, env1, brain1, acting1,
replay_memory1, self._gamma)
agent1.seed(seed)
rewards1, _ = agent1.run()
rewards1 = rewards1.tolist()
env2 = TestEnv()
env2.seed(seed)
brain2 = TestBrain(self._n_states, self._n_actions)
acting2 = EpsGreedyPolicy(self._eps)
replay_memory2 = Memory(self._capacity, self._batch_size)
agent2 = TdAgent(self._n_episodes, env2, brain2, acting2,
replay_memory2, self._gamma)
agent2.seed(seed)
rewards2, _ = agent2.run()
rewards2 = rewards2.tolist()
self.assertListEqual(rewards1, rewards2)
class EpsGreedyPolicyTest(unittest.TestCase):
def setUp(self):
self._eps = 0.1
self._acting = EpsGreedyPolicy(self._eps)
def test_seed(self):
seed = 123
eps = 0.7
lowest_q = 1
highest_q = 10
n_actions = 10
n_qs = 1000
acting1 = EpsGreedyPolicy(eps)
acting1.seed(seed)
acting2 = EpsGreedyPolicy(eps)
acting2.seed(seed)
qs = [
np.random.uniform(lowest_q, highest_q, size=(1, n_actions))
for _ in range(n_qs)
]
for q in qs:
a1 = acting1.act(q)
a2 = acting2.act(q)
self.assertEqual(a1, a2)
def test_act(self):
n_total = 10000
lowest_q = 1
highest_q = 10
n_actions = 100
n_max_q = 0
n_random = 0
for _ in range(n_total):
q = np.random.uniform(lowest_q, highest_q, size=(1, n_actions))
arg_max = np.argmax(q)
action = self._acting.act(q)
if arg_max == action:
n_max_q += 1
else:
n_random += 1
actual = n_random / n_total
max_deviation = 0.1
actual_deviation = abs((self._eps - actual) / self._eps)
self.assertLess(actual_deviation, max_deviation)
def test_seed(self):
seed = 123
eps = 0.7
lowest_q = 1
highest_q = 10
n_actions = 10
n_qs = 1000
acting1 = EpsGreedyPolicy(eps)
acting1.seed(seed)
acting2 = EpsGreedyPolicy(eps)
acting2.seed(seed)
qs = [
np.random.uniform(lowest_q, highest_q, size=(1, n_actions))
for _ in range(n_qs)
]
for q in qs:
a1 = acting1.act(q)
a2 = acting2.act(q)
self.assertEqual(a1, a2)
def setUp(self):
self._n_episodes = 10
self._env = TestEnv()
self._n_states = 3
self._n_actions = 7
self._brain = TestBrain(self._n_states, self._n_actions)
self._eps = 0.1
self._acting = EpsGreedyPolicy(self._eps)
self._capacity = 1000
self._batch_size = 8
self._replay_memory = Memory(self._capacity, self._batch_size)
self._gamma = 0.99
self._train_freq = 4
self._agent = TdAgent(self._n_episodes, self._env, self._brain, self._acting,
self._replay_memory, self._gamma, train_freq=self._train_freq)
def setUp(self):
self._eps = 0.1
self._acting = EpsGreedyPolicy(self._eps)