def test_alphabeta_vs_penalty(self):
t = TTT(3)
player1 = ABPruning(3)
player2 = ABPruning(3)
player2.set_penalty(0.2)
scores = {-4: 0, -2: 0, 0: 0, 1: 0, 3: 0, 5: 0}
game_played = 0
while game_played < 11:
if t.is_terminated():
score = t.get_score()
scores[score] += 1
game_played += 1
t = TTT(3)
pass
mover = t.get_mover()
if mover == 1:
[_, move] = player1.get(t.get_state(), mover)
t.put(move)
elif mover == -1:
[_, move] = player2.get(t.get_state(), mover)
t.put(move)
pass
print(scores)
wrong_cases = scores[-4] + scores[-2]
self.assertTrue(wrong_cases == 0)
def test_diagonal1(self):
t3 = TTT(3)
s3 = np.array([1, 0, -1, 0, 1, -1, 0, 0, 1])
self.assertTrue(t3.is_terminated(s3))
self.assertEqual(t3.check_winner(s3), {
'winner': 1,
'lines': [[0, 4, 8]]
})
t4 = TTT(4)
s4 = np.array([-1, 0, 0, 1, 0, -1, 0, 1, 0, 1, -1, 0, 1, 0, 0, -1])
self.assertTrue(t4.is_terminated(s4))
self.assertEqual(t4.check_winner(s4), {
'winner': -1,
'lines': [[0, 5, 10, 15]]
})
def test_diagonal2(self):
t3 = TTT(3)
s3 = np.array([1, 0, -1, 1, -1, 0, -1, 1, 0])
self.assertEqual(t3.check_winner(s3), {
'winner': -1,
'lines': [[2, 4, 6]]
})
self.assertTrue(t3.is_terminated(s3))
t4 = TTT(4)
s4 = np.array([-1, 0, 0, 1, -1, 0, 1, 0, -1, 1, 0, 0, 1, 0, 0, 0])
self.assertTrue(t4.is_terminated(s4))
self.assertEqual(t4.check_winner(s4), {
'winner': 1,
'lines': [[3, 6, 9, 12]]
})
def __init__(self,size=3):
self._t = TTT(size)
self._mode = 'optimal'
self._penalty_prob = 0
pass
def test_as_second_mover(self):
parameters = {
"ep_train": 0.5,
"ep_infer": 0,
"gamma": 1,
"alpha": 1,
"agent_for": 'minimizer',
}
q = TabularQ(3)
q.set_params(**parameters)
opponent_agent = load('minimax')
q.train(numOfGames=500, opponent_agent=opponent_agent)
t = TTT(3)
Q = q._Q
updated_state_indices = np.where(
Q != [0, 0, 0, 0, 0, 0, 0, 0, 0])[0] # [0] for row indices
updated_state_indices = set(updated_state_indices)
for i in updated_state_indices:
state = q.get_state(i)
mover = t.get_mover(state=state)
self.assertEqual(mover, -1)
return
def test_update(self):
t = TTT(3)
prev_state = [[1, 1, 0], [-1, -1, 0], [0, 0, 0]]
next_state = [[1, 1, 1], [-1, -1, 0], [0, 0, 0]]
prev_state = np.array(prev_state).reshape(-1)
next_state = np.array(next_state).reshape(-1)
result = t.get_result(next_state)
self.assertEqual(result, {'terminated': True, 'score': 5})
q = TabularQ(3)
q.set_params(alpha=1, gamma=1)
encoded_prev_state = t.get_encoded_state(prev_state)
prev_state_index = q.get_index(encoded_prev_state)
encoded_next_state = t.get_encoded_state(next_state)
next_state_index = q.get_index(encoded_next_state)
self.assertEqual(next_state_index, None)
q.update(encoded_prev_state, 2, encoded_next_state, 5)
updated_row = q._Q[prev_state_index, :]
check_row = np.array_equal(updated_row, [0, 0, 5, 0, 0, 0, 0, 0, 0])
self.assertTrue(check_row)
# test correct inference :
q._is_first_mover = True
possible_moves = t.get_available_positions(prev_state)
inferred = q.infer(encoded_prev_state, possible_moves, 1)
self.assertEqual(inferred, 2)
pass
def test_game1(self):
# 0 1 0
# -1 1 0
# -1 1 0
t = TTT(3)
result = t.get_result()
self.assertDictEqual(result, {'terminated': False, 'score': 0})
t.put(1)
result = t.get_result()
self.assertDictEqual(result, {'terminated': False, 'score': 0})
t.put(3)
result = t.get_result()
self.assertDictEqual(result, {'terminated': False, 'score': 0})
t.put(4)
result = t.get_result()
self.assertDictEqual(result, {'terminated': False, 'score': 0})
t.put(6)
result = t.get_result()
self.assertDictEqual(result, {'terminated': False, 'score': 0})
t.put(7)
result = t.get_result()
self.assertDictEqual(result, {'terminated': True, 'score': 5})
return
def test_column(self):
t4 = TTT(4)
s4 = np.array([1, 0, 0, -1, 0, 1, 0, -1, 1, 0, 0, -1, 0, 1, 0, -1])
self.assertEqual(t4.check_winner(s4), {
'winner': -1,
'lines': [[3, 7, 11, 15]]
})
self.assertTrue(t4.is_terminated(s4))
def test_row(self):
t3 = TTT(3)
s3 = np.array([0, 0, 0, -1, -1, 0, 1, 1, 1])
self.assertEqual(t3.check_winner(s3), {
'winner': 1,
'lines': [[6, 7, 8]]
})
self.assertTrue(t3.is_terminated(s3))
def test_get_mover(self):
t = TTT(3)
s = [[0, 1, 0], [0, -1, 0], [0, 0, 0]]
s = np.array(s)
s = s.reshape(-1)
mover = t.get_mover(state=s)
self.assertTrue(mover == 1)
def initialize_minimax(filepath: str, size=3):
table = {}
t = TTT(size)
minimax_save(t.get_state(), t.get_mover(), t, table)
with open(filepath, 'wb') as f:
pickle.dump(table, f)
return
def test_score1(self):
t3 = TTT(3)
s = [[1, -1, 0], [-1, 1, 0], [0, 0, 1]]
s = np.array(s).reshape(-1)
terminated = t3.is_terminated(s)
score = t3.get_score(s)
self.assertTrue(terminated)
self.assertEqual(score, 5)
def test_minimax_1(self):
t = TTT(3)
state = [[1, -1, 0], [-1, 1, 0], [0, 0, 0]]
state = np.array(state, dtype='int')
state = state.reshape(-1)
t._state = state
[score, move] = minimax(t.get_state(), 1, t)
self.assertListEqual(list(state), list(t.get_state()))
self.assertEqual(8, move)
self.assertEqual(5, score)
def run_game(self, agent1, agent2, size=3):
t = TTT(size)
for i in range(size * size):
agent = agent1 if i % 2 == 0 else agent2
inferred = agent(t.get_state())
t.put(inferred)
if t.is_terminated():
break
return t.get_result()
def test_alpha_beta_1(self):
t = TTT(3)
player = ABPruning(3)
state = [[1, -1, 0], [-1, 1, 0], [0, 0, 0]]
state = np.array(state, dtype='int')
state = state.reshape(-1)
t._state = state
[score, move] = player.get(t.get_state(), 1)
self.assertListEqual(list(state), list(t.get_state()))
self.assertEqual(8, move)
self.assertEqual(5, score)
def test_result1(self):
t3 = TTT(3)
s = [[1, -1, -1], [-1, 1, 1], [1, -1, 1]]
s = np.array(s).reshape(-1)
result = t3.get_result(s)
to_equal = {
'terminated': True,
'score': 1,
'winner': 1,
'lines': [[0, 4, 8]]
}
self.assertDictEqual(result, to_equal)
def test_set_state(self):
t = TTT(3)
state = [1, 0, 0, 1, 1, 0, -1, -1, 0]
t.set_state(state)
mover = t.get_mover()
order = t._order
num_of_moves = t._num_moves
_state = np.array(state, dtype=int)
self.assertEqual(mover, -1)
self.assertEqual(order, False)
self.assertEqual(num_of_moves, 5)
self.assertTrue(np.array_equal(_state, t._state))
def test_penalty_vs_penalty(self):
t = TTT(3)
player1 = ABPruning(3)
player1.set_penalty(0.7)
player2 = ABPruning(3)
player2.set_penalty(0.7)
games_played = 1
scores = set()
case1 = {1, 2, 3, 4}
case2 = {-1, -2, -3}
case3 = {0}
while True:
if t.is_terminated():
score = t.get_score()
scores.add(score)
# check whether if win,draw,lose all happened
wins = case1 & scores
loses = case2 & scores
draw = case3 & scores
if len(wins) > 0:
if len(loses) > 0:
if len(draw) > 0:
break
t = TTT(3)
games_played += 1
pass
mover = t.get_mover()
if mover == 1:
[_, move] = player1.get(t.get_state(), mover)
t.put(move)
elif mover == -1:
[_, move] = player2.get(t.get_state(), mover)
t.put(move)
self.assertTrue(len(scores) > 2)
def initialize_state_indices(filepath: str, size=3):
table = {'current': 0} # store state:index pair
t = TTT(size)
def dfs(state, mover: int, table=table) -> None:
# store if the state is new one :
encoded_state = t.get_encoded_state(state)
if not encoded_state in table:
table[encoded_state] = table['current']
table['current'] += 1
assert type(table[encoded_state]) is int
next_mover = 1 if mover is -1 else -1
available_moves = t.get_available_positions(state)
for i in available_moves:
next_state = state.copy()
next_state[i] = mover
if not t.is_terminated(next_state):
dfs(next_state, next_mover)
return
# indexing start :
initial_mover = t.get_mover()
initial_state = t.get_state()
print('indexing start :')
dfs(initial_state, initial_mover)
# simple validate :
num_visited = table['current']
del (table['current'])
num_stored = len(table)
print(f'visited states : {num_visited}')
print(f'stored states : {num_stored}')
assert num_stored == num_visited
indices = set(table.values())
assert len(indices) == len(table)
sample_index = list(table.values())[1]
assert type(sample_index) is int
# save :
print('saving... ', end='')
with open(filepath, 'wb') as f:
pickle.dump(table, f)
print('done')
return
def _train_against(self,opponent_agent:Callable[[np.ndarray],int],numOfGames:int)->None:
agent_q_turn = self._is_first_mover
for _ in tqdm(range(numOfGames)):
game = TTT(self._size)
turn = True
# one complete game :
# prev state, action taken are from agent's turn
# next state is from opponent's turn.
# update in opponent's turn
encoded_prev_state = None
move_taken = None
encoded_next_state = None
while True:
if turn is agent_q_turn:
# Q turn :
if game.is_terminated():
break
else:
possible_moves = game.get_available_positions()
encoded_prev_state = game.get_encoded_state()
move_taken = self._epsilon_greedy_train(encoded_prev_state,possible_moves)
game.put(move_taken)
pass
pass
else:
# opponent's turn :
if not game.is_terminated():
state = game.get_state()
# move below is considered as random (sampling procedure) :
move = opponent_agent(state)
game.put(move)
pass
encoded_next_state = game.get_encoded_state()
score = game.get_score()
if encoded_prev_state is not None:
# : to avoid just after first move case ( in case of Q is second mover )
self.update(encoded_prev_state,move_taken,encoded_next_state,score)
pass
turn = not turn
pass
return None
def test_minimax_vs_minimax(self):
size = 3
t = TTT(size)
filepath = 'results/minimax.pk'
minimax_loaded = minimax_load(filepath)
moves = 0
while True:
[_, best_move] = minimax_loaded(t.get_state())
t.put(best_move)
moves += 1
if t.is_terminated():
break
pass
self.assertEqual(t.check_winner()['winner'], 0)
pass
def test_alphabeta_vs_alphabeta(self):
t = TTT(3)
player = ABPruning(3)
moves = 0
print('Moves : 0 ', end='')
while True:
[_, best_move] = player.get(t.get_state(), t.get_mover())
t.put(best_move)
moves += 1
print(f'{moves} ', end='')
if t.is_terminated():
break
pass
print('final state')
print(t)
self.assertEqual(t.check_winner()['winner'], 0)
def _train_both(self,numOfGames):
for _ in tqdm(range(numOfGames)):
game = TTT(self._size)
self._is_first_mover = True
# one complete game :
while True:
encoded_prev_state = game.get_encoded_state()
possible_moves = game.get_available_positions()
selected_move = self._epsilon_greedy_train(encoded_prev_state,possible_moves)
game.put(selected_move)
encoded_next_state = game.get_encoded_state()
result = game.get_result()
self.update(encoded_prev_state,selected_move,encoded_next_state,result['score'])
if result['terminated']:
break
pass
pass
def __init__(self, user_first: bool, size=3, *args, **kwargs):
super().__init__(*args, **kwargs)
# state variables
self._user_first = user_first
self._t = TTT(size)
self._agent: Callable[[np.ndarray], int]
self._num_of_moves = 0
self._state_history = [self._t.get_state()]
# UI accessors
self._history_scale: tk.Scale
self._player_labels: Dict[int, tk.Label] # key : 1,2
self._buttons = []
# UI initialization
self.title(f'TTT')
self._make_top_frame()
self._make_board(size)
self._make_bottom_frame(size)
return
def test_alphabeta_vs_minimax(self):
t = TTT(3)
minimax_player = minimax_load('results/minimax.pk')
alphabeta_player = ABPruning(3)
moves = 0
print('Moves : 0 ', end='')
while True:
if t.get_mover() == 1:
[_, best_move] = alphabeta_player.get(t.get_state(),
t.get_mover())
elif t.get_mover() == -1:
[_, best_move] = minimax_player(t.get_state())
t.put(best_move)
moves += 1
print(f'{moves} ', end='')
if t.is_terminated():
break
pass
print('final state')
print(t)
self.assertEqual(t.check_winner()['winner'], 0)
def test_deterministic_vs_minimax(self):
# gamma, alpha == 1 guarantees that for endstates s and optimal move a,
# Q(s,a) = R(s,a) IF Q(s,a) IS NOT 0
# Here, R(s,a) is the score of the terminated state
parameters = {
"ep_train": 0.5,
"ep_infer": 0,
"gamma": 1,
"alpha": 1,
"agent_for": 'both',
}
q = TabularQ(3)
q.set_params(**parameters)
q.train(numOfGames=500)
s = Settings()
minimax = minimax_load(s.path('minimax'))
t = TTT(3)
Q = q._Q
to_check_state_indices = np.where(Q != [0, 0, 0, 0, 0, 0, 0, 0, 0])[0]
to_check_state_indices = map(int, to_check_state_indices)
for state_index in to_check_state_indices:
self.assertFalse(
np.array_equal(Q[state_index],
np.array([0, 0, 0, 0, 0, 0, 0, 0, 0])))
state = q.get_state(state_index)
encoded_state = t.get_encoded_state(state)
mover = t.get_mover(state=state)
possible_moves = t.get_available_positions(state)
if mover == 1:
best_move_q = np.argmax(Q[state_index])
if int(Q[state_index, best_move_q]) is not 0:
move_inferred = q.infer(encoded_state, possible_moves,
mover)
q_value_1 = Q[state_index, best_move_q]
q_value_2 = Q[state_index, move_inferred]
self.assertEqual(q_value_1, q_value_2)
elif mover == -1:
best_move_q = np.argmin(Q[state_index])
if int(Q[state_index, best_move_q]) is not 0:
move_inferred = q.infer(encoded_state, possible_moves,
mover)
q_value_1 = Q[state_index, best_move_q]
q_value_2 = Q[state_index, move_inferred]
self.assertEqual(q_value_1, q_value_2)
next_state = state.copy()
next_state[best_move_q] = mover
result = t.get_result(next_state)
if result['terminated']:
best_score, _ = minimax(state)
q_value = Q[state_index, best_move_q]
if best_score != q_value:
# not yet sampled (s,a)
# or withdraw case
self.assertEqual(q_value, 0)
else:
# sampled (s,a)
self.assertEqual(best_score, q_value)
pass
def load(name: str, **kwargs) -> Callable[[np.ndarray], int]:
if name == 'minimax':
from src.utils.path import Settings
s = Settings()
minimax = minimax_load(s.path('minimax'))
def agent(state: np.ndarray) -> int:
move = minimax(state)[1]
return int(move)
return agent
elif name == 'alpha_beta':
assert 'size' in kwargs
assert 'penalty_prob' in kwargs
size = kwargs.get('size')
penalty_prob = kwargs.get('penalty_prob')
ab = ABPruning(size)
ab.set_penalty(penalty_prob)
t = TTT(size)
def agent(state: np.ndarray) -> int:
mover = t.get_mover(state=state)
inferred = ab.get(state=state, mover=mover)[1]
return inferred
return agent
elif name == 'random':
assert 'size' in kwargs
import random
size = kwargs.get('size')
t = TTT(size)
def agent(state: np.ndarray) -> int:
possible_moves = t.get_available_positions(state)
nums = len(possible_moves)
random_index = random.randint(0, nums - 1)
return int(possible_moves[random_index])
return agent
elif name == 'tabular_q':
assert 'id' in kwargs
id = kwargs.get('id')
q = TabularQ.load(id)
size = q._size
t = TTT(size)
def agent(state: np.ndarray) -> int:
possible_moves = t.get_available_positions(state)
encoded_state = t.get_encoded_state(state)
mover = t.get_mover(state=state)
inferred = q.infer(encoded_state, possible_moves, mover)
return inferred
return agent
else:
raise NameError(f'{name} is not implemented')
def test_availables(self):
t3 = TTT(3)
s3 = [[1, -1, 0], [0, 1, -1], [1, -1, 0]]
s3 = np.array(s3).reshape(-1)
indices = t3.get_available_positions(s3)
self.assertListEqual(indices, [2, 3, 8])
def test_state_encode(self):
t3 = TTT(3)
encoded01 = t3.get_encoded_state()
state02 = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0], dtype=int)
encoded02 = t3.get_encoded_state(state02)
self.assertEqual(encoded01, encoded02)