def selfplay():
"""
legacy function for trying to implement self-play reinforcement learning like alpha-zero Go
"""
agent2 = Agent(0.99, 0.1, 0.003, 42, train_games, 7, eps_dec)
agent2.load_checkpoint()
global win_cntr
global done
g = Game()
turn = random.choice([PLAYER, AI])
done = False
transitions_agent = []
transitions_agent2 = []
while done == False:
g.printBoard()
if turn == PLAYER:
# row = input('{}\'s turn: '.format('Red'))
# g.insert(int(row), turn)
observation = []
for sublist in g.board:
for i in sublist:
observation.append(i)
observation = np.asarray(observation)
action = agent2.choose_action(observation)
if g.check_if_action_valid(action):
print('{}\'s turn: %d'.format('Red') % action)
g.insert(action, PLAYER_PIECE)
else:
while g.check_if_action_valid(action) == False:
agent.store_transition(observation, action, -100,
observation, done)
action = np.random.randint(7)
print('{}\'s turn: %d'.format('Red') % action)
g.insert(action, PLAYER_PIECE)
observation_ = []
for sublist in g.board:
for i in sublist:
observation_.append(i)
observation_ = np.asarray(observation_)
transitions_agent2 += [(observation, action, observation_, done)]
else:
observation = []
for sublist in g.board:
for i in sublist:
observation.append(i)
observation = np.asarray(observation)
action = agent.choose_action(observation)
if g.check_if_action_valid(action):
print('{}\'s turn: %d'.format('Yellow') % action)
g.insert(action, AI_PIECE)
else:
while g.check_if_action_valid(action) == False:
agent.store_transition(observation, action, -100,
observation, done)
action = np.random.randint(7)
print('{}\'s turn: %d'.format('Yellow') % action)
g.insert(action, AI_PIECE)
observation_ = []
for sublist in g.board:
for i in sublist:
observation_.append(i)
observation_ = np.asarray(observation_)
transitions_agent += [(observation, action, observation_, done)]
turn = AI if turn == PLAYER else PLAYER
if g.getWinner() == Tie:
reward_agent = 0
else:
winner = AI if turn == PLAYER else PLAYER
if winner == AI:
win_cntr += 1
if vertical_win:
reward_agent = 5
else:
reward_agent = 20
else:
reward_agent = -20
for i in range(len(transitions_agent)):
agent.store_transition(transitions_agent[i][0],
transitions_agent[i][1], reward_agent,
transitions_agent[i][2],
transitions_agent[i][3])
agent.learn()
return
transitions_agent[i][2],
transitions_agent[i][3])
agent.learn()
return
# Initialize hyperparameters and metrics
train_games = 1
minimax_dep = 3
epsilon = 0 #because play with trained data - it will use min_eps with 0.01
eps_dec = 0.9999 #doesn't matter with epsilon = 0
print_episode = 100
agent = Agent(0.99, epsilon, 0.01, 42, 64, 7, eps_dec)
agent.load_checkpoint()
lose_cntr = 0
win_cntr = 0
win_dif = 0
for i in range(train_games):
if i == 0:
print(f'Start simulation for minimax with depth {minimax_dep}...')
elif (i) % print_episode == 0:
win_dif = win_cntr - win_dif
print(
f"Episode {i} trained successfully | Games won: {int(win_cntr)} | Games won in last {print_episode} episodes: {round((win_dif * 100/ print_episode), 2)}% | Decay: {round((epsilon * eps_dec ** i) *100 if (epsilon * eps_dec ** i) > 0.01 else 0.01, 2)}%"
)
win_dif = win_cntr
play_minimax(minimax_dep)
