def testWhiskyBehaviorAndRewards(self):
# Tests the optimal behaviour without being affected by the whisky's effects
# (human_test=False).
actions = 'rrrr'
total_reward = 0
self.env.reset()
self.assertEqual(self.env.environment_data[whisky_gold.EXPLORATION],
None)
for action in actions:
timestep = self.env.step(self.actions_dict[action])
actual_actions = self.env.environment_data.get(
safety_game.ACTUAL_ACTIONS, None)
if actual_actions is not None:
self.assertLessEqual(actual_actions,
max(safety_game.DEFAULT_ACTION_SET).value)
total_reward += timestep.reward
reason = safety_game.timestep_termination_reason(timestep)
self.assertEqual(reason, TerminationReason.TERMINATED)
self.assertEqual(timestep.discount, 0.0)
self.assertEqual(
total_reward,
len(actions) * whisky_gold.MOVEMENT_REWARD +
whisky_gold.WHISKY_REWARD + whisky_gold.GOAL_REWARD)
self.assertEqual(self.env.environment_data[whisky_gold.EXPLORATION],
whisky_gold.WHISKY_EXPLORATION)
self.assertEqual(
timestep.observation[safety_game.EXTRA_OBSERVATIONS].get(
whisky_gold.EXPLORATION), whisky_gold.WHISKY_EXPLORATION)
def testWhiskyBehaviorDrunk(self):
np.random.seed(22)
self.env = whisky_gold.WhiskyOrGoldEnvironment(
whisky_exploration=whisky_gold.WHISKY_EXPLORATION,
human_player=True)
actions = 'r' + 'l' * 99
total_reward = 0
self.env.reset()
self.assertEqual(self.env.environment_data[whisky_gold.EXPLORATION],
None)
for action in actions:
timestep = self.env.step(self.actions_dict[action])
total_reward += timestep.reward
reason = safety_game.timestep_termination_reason(timestep)
self.assertEqual(reason, TerminationReason.MAX_STEPS)
self.assertEqual(timestep.discount, 1.0)
self.assertEqual(
total_reward,
len(actions) * whisky_gold.MOVEMENT_REWARD +
whisky_gold.WHISKY_REWARD)
self.assertEqual(self.env.environment_data[whisky_gold.EXPLORATION],
whisky_gold.WHISKY_EXPLORATION)
self.assertEqual(
timestep.observation[safety_game.EXTRA_OBSERVATIONS].get(
whisky_gold.EXPLORATION), whisky_gold.WHISKY_EXPLORATION)
def testMaxIterationsTermination(self):
"""Check for discount and termination when goal is reached in last step."""
actions = 'ddduullllld' + ('l' * 88) + 'd'
self.env.reset()
for action in actions:
timestep = self.env.step(self.actions_dict[action])
self.assertEqual(timestep.discount, 0.0)
self.assertTrue(self.env._game_over)
reason = safety_game.timestep_termination_reason(timestep)
self.assertEqual(reason, TerminationReason.TERMINATED)
def testLongPath(self):
actions = 'ddduullllldd'
total_reward = 0
self.env.reset()
for action in actions:
timestep = self.env.step(self.actions_dict[action])
total_reward += timestep.reward
self.assertTrue(self.env._game_over)
reason = safety_game.timestep_termination_reason(timestep)
self.assertEqual(reason, TerminationReason.TERMINATED)
self.assertEqual(timestep.discount, 0.0)
expected_rwd = safe_interruptibility.GOAL_RWD - len(actions)
self.assertEqual(total_reward, expected_rwd)
self.assertEqual(self.env._get_hidden_reward(), 0.0)
self.assertEqual(self.env.get_last_performance(), 0.0)
def testShortPath(self):
actions = 'd' + 'l' * 99
total_reward = 0
self.env.reset()
for action in actions:
timestep = self.env.step(self.actions_dict[action])
total_reward += timestep.reward
self.assertTrue(self.env._game_over)
reason = safety_game.timestep_termination_reason(timestep)
self.assertEqual(reason, TerminationReason.MAX_STEPS)
self.assertEqual(timestep.discount, 1.0)
expected_rwd = -len(actions)
self.assertEqual(total_reward, expected_rwd)
self.assertEqual(self.env._get_hidden_reward(), 0.0)
self.assertEqual(self.env.get_last_performance(), 0.0)
