def play(strategy, verbose=False, allow_unavailable_action=True):
"""Plays a single game, using a provided strategy.
Args:
strategy: A function that takes as argument a state and a list of available
actions and returns an action from the list.
allow_unavailable_action: Boolean, whether strategy is passed all actions
or just the available ones.
verbose: If true, prints game states, actions and scores.
Returns:
score, experiences where score is the final score and experiences is the
list Experience instances that represent the collected experience.
"""
game = Game()
state = game.state().copy()
game_over = game.game_over()
experiences = []
while not game_over:
if verbose:
print("Score:", game.score())
game.print_state()
old_state = state
next_action = strategy(
old_state,
range(4) if allow_unavailable_action else game.available_actions())
if game.is_action_available(next_action):
reward = game.do_action(next_action)
state = game.state().copy()
game_over = game.game_over()
if verbose:
print("Action:", ACTION_NAMES[next_action])
print("Reward:", reward)
experiences.append(
Experience(old_state, next_action, reward, state, game_over,
False, game.available_actions()))
else:
experiences.append(
Experience(state, next_action, 0, state, False, True,
game.available_actions()))
if verbose:
print("Score:", game.score())
game.print_state()
print("Game over.")
return game.score(), experiences
def test_available_actions():
state = np.array([[[1, 2, 0], [1, 2, 0], [1, 2, 0]],
[[1, 2, 0], [1, 2, 0], [1, 2, 0]],
[[1, 2, 0], [1, 2, 0], [1, 2, 0]]])
game = Game(state=state)
actions = game.available_actions()
# All actions except left is available
assert actions == [1, 2, 3, 4, 5]
def test_available_actions_none_available():
state = np.array([[[1, 2, 3], [5, 6, 7], [1, 2, 3]],
[[8, 9, 10], [11, 12, 13], [8, 9, 10]],
[[14, 15, 16], [17, 18, 19], [14, 15, 16]]])
game = Game(state=state)
actions = game.available_actions()
# All actions except left is available
assert actions == []
assert game.game_over()
def highest_reward_strategy(state, actions):
"""Strategy that always chooses the action of highest immediate reward.
If there are any ties, the strategy prefers left over up over right over down.
"""
sorted_actions = np.sort(actions)[::-1]
rewards = map(lambda action: Game(np.copy(state)).do_action(action),
sorted_actions)
action_index = np.argsort(rewards, kind="mergesort")[-1]
return sorted_actions[action_index]
def test_available_actions_none_available():
state = np.array([[1, 2, 3, 4],
[5, 6, 7, 8],
[1, 2, 3, 4],
[5, 6, 7, 8]])
game = Game(state=state)
actions = game.available_actions()
# All actions except left is available
assert actions == []
assert game.game_over()
def highest_reward_strategy(state, actions):
"""Strategy that always chooses the action of highest immediate reward.
If there are any ties, the strategy prefers left over up over right over down.
"""
sorted_actions = np.sort(actions)[::-1]
rewards = [
Game(np.copy(state)).do_action(action) for action in sorted_actions
]
action_index = np.argsort(rewards)[-1]
return sorted_actions[action_index]
def test_init(choice):
choice.side_effect = [0, # First position
1, # First tile
1, # Second position
2] # Second tile
game = Game()
choice.assert_has_calls([call(16),
call([1, 2], p=[0.9, 0.1]),
call(15),
call([1, 2], p=[0.9, 0.1])])
# Assert correct number of 0s, 1s and 2s
game.print_state()
assert (np.bincount(game.state().flatten()) == [14, 1, 1]).all()
assert game.score() == 0
def test_do_action(choice):
choice.side_effect = [0, # First position
1] # First tile
state = np.array([[1, 2, 3, 3],
[5, 6, 7, 8],
[5, 2, 7, 0],
[1, 0, 3, 0]])
game = Game(state=state)
game.do_action(3) # DOWN
new_state = np.array([[1, 0, 0, 0],
[1, 2, 3, 0],
[6, 6, 8, 3],
[1, 2, 3, 8]])
game.print_state()
assert (game.state() == new_state).all()
# Score is 2 ** 6 + 2 ** 8
assert game.score() == 320
def test_do_action(choice):
choice.side_effect = [
0, # First position
1
] # First tile
state = np.array([[[1, 2, 3], [5, 6, 7], [5, 2, 7]],
[[1, 2, 3], [5, 6, 7], [5, 2, 7]],
[[1, 2, 3], [5, 6, 7], [5, 2, 7]]])
game = Game(state=state)
game.do_action(3) # DOWN
new_state = np.array([[[0, 2, 0], [1, 6, 3], [6, 2, 8]],
[[0, 2, 0], [1, 6, 3], [6, 2, 8]],
[[0, 2, 0], [1, 6, 3], [6, 2, 8]]])
game.print_state()
assert (game.state() == new_state).all()
# Score is (2 ** 6 + 2 ** 8)*3
assert game.score() == 960
target_f = self.model.predict(state)
target_f[0][action] = target
self.model.fit(state, target_f, epochs=1, verbose=0)
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay
def load(self, name):
self.model.load_weights(name)
def save(self, name):
self.model.save_weights(name)
if __name__ == "__main__":
count = 0
game = Game()
env = gym.make('CartPole-v1')
#state_size = 192
state_size = env.observation_space.shape[0] # in our case 4*4*12
#action_size = 4
action_size = env.action_space.n #in our case 4
agent = DQNAgent(state_size, action_size)
# agent.load("./save/cartpole-dqn.h5")
done = False
batch_size = 32
for e in range(EPISODES):
state = game.state()
count += 1
#state = env.reset()
#state = np.reshape(state, [1, state_size])
