def deep_q_learning_step(epsilon, player):
global loss_for_one_episode
index = epsilon_greedy(epsilon, player)
q_value = (model(torch.FloatTensor(game.board))[(player + 2) % 3])[index]
a_p, reward = game.step(index, player)
if abs(a_p) == 10 or game.full_board():
loss = ((reward - q_value)**2)
else:
while a_p != player and abs(a_p) != 10 and not game.full_board():
index = epsilon_greedy(agr, a_p)
a_p, _ = game.step(index, a_p)
if abs(a_p) == 10:
loss = ((reward - 17 - q_value)**2)
elif game.full_board():
loss = ((reward - 5 - q_value)**2)
else:
q_value_max = (model(torch.FloatTensor(game.board) *
player)[(a_p + 2) % 3]).max()
loss = ((reward + GAMMA * q_value_max - q_value)**2)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_for_one_episode = loss_for_one_episode + loss
return a_p
def play_with():
game.new_game()
player = 1
print(game.board)
while abs(player) != 10 and not game.full_board():
index = epsilon_greedy(0.0, player)
player, _ = game.step(index, player)
print(game.board)
if not (abs(player) != 10 and not game.full_board()):
continue
my_index = -1 + int(input("index: "))
player, _ = game.step(my_index, player)
print(game.board)
def one_episode(epsilon, player):
game.new_game()
global loss_for_one_episode, loss_for_sever_episodes
loss_for_one_episode = 0
if player == 1:
while abs(player) != 10 and not game.full_board():
player = deep_q_learning_step(epsilon, player)
else:
index = epsilon_greedy(0.0, 1)
player, _ = game.step(index, player)
while abs(player) != 10 and not game.full_board():
player = deep_q_learning_step(epsilon, player)
print(loss_for_one_episode)
loss_for_sever_episodes += loss_for_one_episode
def test():
game.new_game()
player = 1
print(game.board)
while abs(player) != 10 and not game.full_board():
index = epsilon_greedy(0.0, player)
player, _ = game.step(index, player)
print(game.board)