# Start Play state
if b_:
gamestate = PLAY_STATE
ttt = Tictactoe()
time.sleep(0.1)
# Play state
elif gamestate == PLAY_STATE:
# Mouse click
if event.type == pygame.MOUSEBUTTONDOWN:
point = event.pos
# If user's trun
if not ai_turn:
# If game is not over
if not ttt.gameover():
# Cells
b_ = True
for i in range(3):
for j in range(3):
if cells[i][j].collidepoint(point):
if ttt.is_valid_action((i, j)):
ttt.action((i, j))
ai_turn = True
break
# Menu button
if menu_button.collidepoint(point):
gamestate = MENU_STATE
def train(self, num_games):
"""
Trains the agent by playing games against itself
Args:
num_games (int): number of games to train
"""
# Play num_games games
for n in range(num_games):
# Print game number
if n % 1000 == 0:
print(f'Game #{n + 1}')
# Initialize the game
ttt = Tictactoe()
# Keep track of last state and actions
last = {
'X': {
'state': None,
'action': None
},
'O': {
'state': None,
'action': None
}
}
# Play the game
while True:
# Get the state and action
state = ttt.get_board()
action = self.best_action(state, epsilon_true=True)
# Save as lasts
last[ttt.get_player()]['state'] = state
last[ttt.get_player()]['action'] = action
# Apply action and get the new state
ttt.action(action)
new_state = ttt.get_board()
# Game over
if ttt.gameover():
# Won the game
if ttt.get_winner() is not None:
# Update q value for winner
self.update_q_value(state, action, new_state, 1)
# Update q value for loser
self.update_q_value(last[ttt.get_player()]['state'],
last[ttt.get_player()]['action'],
new_state, -1)
# Draw
else:
# Update q value
self.update_q_value(state, action, new_state, 0)
break
# Game continues
elif last[ttt.get_player()]['state'] is not None:
# Update last action
self.update_q_value(last[ttt.get_player()]['state'],
last[ttt.get_player()]['action'],
new_state, 0)
print('Training done')