def play(show_number):
env = TicTacToeEnv(show_number=show_number)
agents = [MinimaxAgent('O'),
HumanAgent('X')]
episode = 0
while True:
state = env.reset()
_, mark = state
done = False
env.render()
while not done:
agent = agent_by_mark(agents, mark)
env.show_turn(True, mark)
ava_actions = env.available_actions()
if mark=='O':
n,action=agent.act(state, ava_actions)
else:
action = agent.act(state, ava_actions)
if action is None:
sys.exit()
state, reward, done, info = env.step(action)
print('')
env.render()
if done:
env.show_result(True, mark, reward)
break
else:
_, _ = state
mark = next_mark(mark)
episode += 1
def play(max_episode=10):
episode = 0
start_mark = 'O'
env = TicTacToeEnv()
agents = [BaseAgent('O'), BaseAgent('X')]
while episode < max_episode:
env.set_start_mark(start_mark)
state = env.reset()
_, mark = state
done = False
while not done:
env.show_turn(True, mark)
agent = agent_by_mark(agents, mark)
ava_actions = env.available_actions()
action = agent.act(state, ava_actions)
state, reward, done, info = env.step(action)
env.render()
if done:
env.show_result(True, mark, reward)
break
else:
_, mark = state
# rotate start
start_mark = next_mark(start_mark)
episode += 1
def learn_on_policy(episodes, epsilon=0.1, discount_factor=0.9):
env = TicTacToeEnv()
agents = [MCOPA('O', epsilon), MCOPA('X', epsilon)]
start_mark = 'O'
env.set_start_mark(start_mark)
for i in range(episodes):
episode = i + 1
state = env.reset()
_, mark = state
steps = []
done = False
while not done:
agent = agent_by_mark(agents, mark)
actions = env.available_actions()
action = agent.act(state, actions)
next_state, reward, done, _ = env.step(action)
steps.append((state, reward))
if done:
break
_, mark = state = next_state
steps.reverse()
G = 0
# As in one episode of tic tac toe there will only be unique states we don't need to check for them
for step in steps:
_, mark = step[0]
G = step[1] + discount_factor * G
agents[0].update(step[0], G)
# rotate start
start_mark = next_mark(start_mark)
def _bench(max_episode, model_file, show_result=True):
"""Benchmark given model.
Args:
max_episode (int): Episode count to benchmark.
model_file (str): Learned model file name to benchmark.
show_result (bool): Output result to stdout.
Returns:
(dict): Benchmark result.
"""
minfo = load_model(model_file)
agents = [BaseAgent('O'), TDAgent('X', 0, 0)]
show = False
start_mark = 'O'
env = TicTacToeEnv()
env.set_start_mark(start_mark)
episode = 0
results = []
for i in tqdm(range(max_episode)):
env.set_start_mark(start_mark)
state = env.reset()
_, mark = state
done = False
while not done:
agent = agent_by_mark(agents, mark)
ava_actions = env.available_actions()
action = agent.act(state, ava_actions)
state, reward, done, info = env.step(action)
if show:
env.show_turn(True, mark)
env.render(mode='human')
if done:
if show:
env.show_result(True, mark, reward)
results.append(reward)
break
else:
_, mark = state
# rotation start
start_mark = next_mark(start_mark)
episode += 1
o_win = results.count(1)
x_win = results.count(-1)
draw = len(results) - o_win - x_win
mfile = model_file.replace(CWD + os.sep, '')
minfo.update(
dict(base_win=o_win, td_win=x_win, draw=draw, model_file=mfile))
result = json.dumps(minfo)
if show_result:
print(result)
return result
def _play(load_file, vs_agent, show_number):
"""Play with learned model.
Make TD agent and adversarial agnet to play with.
Play and switch starting mark when the game finished.
TD agent behave no exploring action while in play mode.
Args:
load_file (str):
vs_agent (object): Enemy agent of TD agent.
show_number (bool): Whether show grid number for visual hint.
"""
load_model(load_file)
env = TicTacToeEnv(show_number=show_number)
td_agent = TDAgent('X', 0, 0) # prevent exploring
start_mark = 'O'
agents = [vs_agent, td_agent]
while True:
# start agent rotation
env.set_start_mark(start_mark)
state = env.reset()
_, mark = state
done = False
# show start board for human agent
if mark == 'O':
env.render(mode='human')
while not done:
agent = agent_by_mark(agents, mark)
human = isinstance(agent, HumanAgent)
env.show_turn(True, mark)
ava_actions = env.available_actions()
if human:
action = agent.act(ava_actions)
if action is None:
sys.exit()
else:
action = agent.act(state, ava_actions)
state, reward, done, info = env.step(action)
env.render(mode='human')
if done:
env.show_result(True, mark, reward)
break
else:
_, mark = state
# rotation start
start_mark = next_mark(start_mark)
def learn_off_policy(episodes, discount_factor=0.9):
env = TicTacToeEnv()
agents = [MCOffPA('O'),
MCOffPA('X')]
start_mark = 'O'
env.set_start_mark(start_mark)
for i in range(episodes):
state = env.reset()
_, mark = state
steps = []
done = False
while not done:
agent = agent_by_mark(agents, mark)
actions = env.available_actions()
action = random.choice(actions)
next_state, reward, done, _ = env.step(action)
steps.append((state, reward, action, actions))
if done:
break
_, mark = state = next_state
steps.reverse()
G = 0
W = 1
# As in one episode of tic tac toe there will only be unique states we don't need to check for them
for step in steps:
_, mark = step[0]
agent = agent_by_mark(agents, mark)
G = step[1] + discount_factor*G
agent.update(step[0], G, W)
if agent.act(step[0], step[3]) != step[2]:
break
# behaviour policy = 1/available_actions
W = W*len(step[3])
# rotate start
start_mark = next_mark(start_mark)
def _learn(max_episode, epsilon, alpha, save_file):
"""Learn by episodes.
Make two TD agent, and repeat self play for given episode count.
Update state values as reward coming from the environment.
Args:
max_episode (int): Episode count.
epsilon (float): Probability of exploration.
alpha (float): Step size.
save_file: File name to save result.
"""
reset_state_values()
env = TicTacToeEnv()
agents = [TDAgent('O', epsilon, alpha), TDAgent('X', epsilon, alpha)]
start_mark = 'O'
for i in tqdm(range(max_episode)):
episode = i + 1
env.show_episode(False, episode)
# reset agent for new episode
for agent in agents:
agent.episode_rate = episode / float(max_episode)
env.set_start_mark(start_mark)
state = env.reset()
_, mark = state
done = False
while not done:
agent = agent_by_mark(agents, mark)
ava_actions = env.available_actions()
env.show_turn(False, mark)
action = agent.act(state, ava_actions)
# update (no rendering)
nstate, reward, done, info = env.step(action)
agent.backup(state, nstate, reward)
if done:
env.show_result(False, mark, reward)
# set terminal state value
set_state_value(state, reward)
_, mark = state = nstate
# rotate start
start_mark = next_mark(start_mark)
# save states
save_model(save_file, max_episode, epsilon, alpha)
def plotting_MC(env, agent1, agent2, learn_fn, r=100):
num_states = []
num_roll_out = []
ub = [[], []]
lb = [[], []]
mean = [[], []]
agent1.reset()
# Training curves
k = 0
for j in range(r):
num_roll_out.append(k)
num_states.append(agent1.num_states())
for i in range(15):
R = 0
for _ in range(10):
state = env.reset()
_, mark = state
done = False
agents = [agent1, agent2]
while not done:
agent = agent_by_mark(agents, mark)
actions = env.available_actions()
action = agent.act(state, actions)
next_state, reward, done, _ = env.step(action)
_, mark = state = next_state
R += reward
if len(ub[0]) == j:
ub[0].append(R)
lb[0].append(R)
mean[0].append(R)
ub[1].append(-R)
lb[1].append(-R)
mean[1].append(-R)
else:
ub[0][j] = max(ub[0][j], R)
lb[0][j] = min(lb[0][j], R)
mean[0][j] = mean[0][j] + (R - mean[0][j]) / (i + 1)
ub[1][j] = max(ub[1][j], -R)
lb[1][j] = min(lb[1][j], -R)
mean[1][j] = mean[1][j] + (-R - mean[1][j]) / (i + 1)
learn_fn(500)
k += 500
# plot the data
fig = plt.figure(1, figsize=(14, 5))
plot_mean_and_CI(mean[0],
ub[0],
lb[0],
num_states,
color_mean='r',
color_shading='r')
plot_mean_and_CI(mean[1],
ub[1],
lb[1],
num_states,
color_mean='b',
color_shading='b')
plt.title('Confidence interval vs states covered')
plt.show()
# plot the data
fig = plt.figure(2, figsize=(14, 5))
plot_mean_and_CI(mean[0],
ub[0],
lb[0],
num_roll_out,
color_mean='r',
color_shading='r')
plot_mean_and_CI(mean[1],
ub[1],
lb[1],
num_roll_out,
color_mean='b',
color_shading='b')
plt.title('Confidence interval vs number of rollouts')
plt.show()