def test_stick_20_policy_save(self):
game = Easy21()
policy = Stick20ActionPolicy(game.action_space)
for time_steps in [1_000]:
mc = MonteCarloPolicyEvaluation(env=game, policy=policy)
mc.learn(total_timesteps=time_steps)
mc.save("stick20_%s" % time_steps)
def test_game_reset(self):
game = Easy21()
self.assertEqual(game.score_dealer, 0)
self.assertEqual(game.score_player, 0)
game.reset()
self.assertNotEqual(game.score_dealer, 0)
self.assertNotEqual(game.score_player, 0)
def test_game_step_stick(self):
game = Easy21()
game.reset()
(dealer, player), reward, done, _ = game.step(action=0)
self.assertEqual(done, True, msg="Done should be 'True'")
self.assertIn(reward, [-1, 0, 1],
msg="Rewards should be in [-1, 0, 1]")
self.__assert_scores(dealer, player, reward)
def test_epsilon_greedy__save(self):
for _lambda in [.0, .1, .2, .3, .4, .5, .6, .7, .8, .9, 1]:
print("Training with lambda: %s" % _lambda)
game = Easy21()
policy = EpsilonGreedyActionPolicy(game.action_space)
model = Sarsa(env=game, policy=policy, _lambda=_lambda)
model.learn(total_timesteps=1000)
model.save("greedy_1000_%s" % _lambda)
def test_random_policy(self):
game = Easy21()
policy = RandomActionPolicy(game.action_space)
for time_steps in [1_000, 10_000, 100_000]:
mc = MonteCarloPolicyEvaluation(env=game, policy=policy)
mc.learn(total_timesteps=time_steps)
game_plots.plot(lambda d, p: mc.q_value(observation=(d, p), action=ACTION_STICK))
game_plots.plot(lambda d, p: mc.q_value(observation=(d, p), action=ACTION_HIT))
game_plots.plot(lambda d, p: mc.q_value_max(observation=(d, p)))
game_plots.plot_image(lambda d, p: mc.q_max(observation=(d, p)))
def test_game_step_hit(self):
game = Easy21()
game.reset()
dealer_before = game.score_dealer
player_before = game.score_player
(dealer, player), reward, done, _ = game.step(action=1)
self.assertEqual(done, False)
self.assertEqual(reward, 0)
self.assertEqual(dealer_before, dealer)
self.assertNotEqual(player_before, player)
def test_game(self):
env = Easy21()
env.reset()
while True:
env.render()
(dealer, player), reward, done, _ = env.step(
action=env.action_space.sample())
if done:
break
env.render()
self.assertIn(reward, [-1, 0, 1])
self.__assert_scores(dealer, player, reward)
def test_epsilon_greedy_policy_1M_save(self):
game = Easy21()
policy = EpsilonGreedyActionPolicy(game.action_space)
mc = MonteCarloPolicyEvaluation(env=game, policy=policy)
mc.learn(total_timesteps=1_000_000)
mc.save("greedy1M")