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 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_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_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 _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 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
class GameWindow(tk.Toplevel):
"""Game UI"""
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
#region Public Methods
def set_agent(self, agent: Callable[[np.ndarray], int], name: str) -> None:
self._agent = agent
return
def get_result(self) -> dict:
return self._t.get_result()
#endregion
#region Put UI Components
def _make_top_frame(self):
frame = tk.Frame(self)
if self._user_first:
text1 = 'O : User'
text2 = 'X : AI'
else:
text1 = 'O : AI'
text2 = 'X : User'
label1 = tk.Label(frame, text=text1)
label2 = tk.Label(frame, text=text2)
label1.pack()
label2.pack()
frame.pack()
return
def _make_board(self, size):
board = tk.Frame(self)
buttons = self._buttons
num_of_buttons = size * size
for i in range(num_of_buttons):
b = tk.Button(board,
width=3,
height=1,
font=('Helvetica', 30),
activebackground='white',
command=lambda num=i: self._on_click_board(num))
buttons.append(b)
b.grid(column=i % size, row=int(i / size))
pass
board.pack()
return
def _make_bottom_frame(self, size):
frame = tk.Frame(self)
history_scale = tk.Scale(frame,
command=self._on_scale_move,
orient='horizontal',
from_=0,
to=0)
history_scale.grid(row=0, columnspan=2)
self._history_scale = history_scale
restart_button = tk.Button(frame,
text="Restart",
command=self._on_click_reset)
exit_button = tk.Button(frame, text="Exit", command=self.destroy)
restart_button.grid(row=1, column=0)
exit_button.grid(row=1, column=1)
frame.pack()
return
#endregion
#region Event Handlers
def _on_click_board(self, position: int):
state_num = int(self._history_scale.get())
is_rewinded = not (self._num_of_moves == state_num)
if is_rewinded:
# reset the game to the rewinded one :
state_to_force = self._state_history[state_num]
self._t.set_state(state_to_force)
self._num_of_moves = self._t._num_moves
self._state_history = self._state_history[0:(self._num_of_moves +
1)]
pass
self._t.put(position)
current_state = self._t.get_state()
self._state_history.append(current_state)
self._num_of_moves += 1
self._history_scale.configure(to=self._num_of_moves)
self._history_scale.set(self._num_of_moves)
"""
[issue] If this procedure is called by button.invoke()
then it doesn't invoke the scale's command _on_scale_move.
So call it manually (and hence, called twice in user's turn) :
"""
self._on_scale_move(self._num_of_moves)
return
def _on_scale_move(self, state_num):
state_num = int(state_num)
first_mover_turn = True if state_num % 2 == 0 else False
user_turn = first_mover_turn == self._user_first
self._set_board(state_num, user_turn)
if self.get_result()['terminated']:
return
if state_num == len(self._state_history) - 1:
if user_turn:
pass
else:
if hasattr(self, '_agent'):
self._on_agent_turn(state_num)
pass
else:
# : agent's turn but it's a previous state
pass
return
def _on_click_reset(self):
self._num_of_moves = 0
self._state_history = self._state_history[0:1]
self._t.set_state(self._state_history[0])
self._history_scale.configure(to=0)
self._history_scale.set(0)
self._set_board(0, self._user_first == True)
return
#endregion
#region Private Methods
def _on_agent_turn(self, state_num: int):
# TODO : async progress bar
state = self._state_history[state_num]
move = self._agent(state)
button = self._buttons[move]
button.configure(state='normal')
button.invoke()
return
def _set_board(self, state_num: int, user_turn: bool):
"""Modify board UI"""
to_state = self._state_history[state_num]
result = self._t.get_result(to_state)
terminated = result['terminated']
lines = result['lines']
lines = sum(lines, []) # flattening
for p in range(len(to_state)):
move = int(to_state[p])
of_line = p in lines
self._modify_button(p, move, user_turn, terminated, of_line)
return
def _modify_button(self,
button_position: int,
mover: int,
move_allowed: bool,
terminated=False,
of_line=False):
button = self._buttons[button_position]
args = {'disabledforeground': 'black', 'state': 'disabled'}
if mover == 1:
args['text'] = '○'
args['state'] = 'disabled'
elif mover == -1:
args['text'] = '×'
args['state'] = 'disabled'
else:
args['text'] = ' '
if move_allowed:
args['state'] = 'normal'
elif not hasattr(self, '_agent'):
args['state'] = 'normal'
if terminated:
args['state'] = 'disabled'
if of_line:
if mover == 1:
args['disabledforeground'] = 'steelblue'
elif mover == -1:
args['disabledforeground'] = 'tomato'
button.config(**args)
return