def test_rush_policy():
"""
Only one possible move.
"""
state = State(board=[[1, 2, 1], [2, 2, 1], [0, 1, 2]], turn=1)
policy = TabularPolicy()
after_state = State(from_base10=policy.move_dict[state.get_num()])
expected_after_state = State(board=[[1, 2, 1], [2, 2, 1], [1, 1, 2]],
turn=2)
assert after_state.board == expected_after_state.board
assert after_state.turn == expected_after_state.turn
"""
Multiple possible moves.
"""
state = State(board=[[1, 0, 0], [2, 2, 1], [0, 1, 2]], turn=2)
policy = TabularPolicy()
after_state = State(from_base10=policy.move_dict[state.get_num()])
expected_board = [[1, 2, 0], [2, 2, 1], [0, 1, 2]]
assert after_state.board == expected_board
assert after_state.turn == 1
"""
Filled board
"""
state = State(board=[[1, 2, 1], [2, 2, 1], [1, 1, 2]], turn=2)
policy = TabularPolicy()
with pytest.raises(KeyError):
after_state = State(from_base10=policy.move_dict[state.get_num()])
def test_random_move():
policy = TabularPolicy(epsilon=1)
collect = []
for _ in range(10000):
collect.append(policy.move(31206))
assert collect.count(51618) == pytest.approx(3333, abs=100)
assert collect.count(50916) == pytest.approx(3333, abs=100)
assert collect.count(50890) == pytest.approx(3333, abs=100)
def test_be_greedy():
policy = TabularPolicy()
best = State(board=[[0, 0, 0], [1, 0, 0], [0, 0, 0]], turn=2)
policy.v_dict[best.get_num()] = 1
assert policy.be_greedy()
state = State()
assert policy.move_dict[state.get_num()] == best.get_num()
assert not policy.be_greedy() # No more change when run the second time
class Train:
def __init__(self, path, read_first=False):
"""
Input:
path: the path to save the policy
read_first: if true, read from the path first
"""
if read_first:
self.target_policy, self.i_epoch = pickle.load(open(path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.target_policy = TabularPolicy()
self.i_epoch = 0
self.opponent_policy = TabularPolicy()
self.path = path
# num for the state with an empty board and with player 1 to make a move.
self.start_num = int('1' + '0' * 9, 3)
def TrainContinuously(self, n_epoch=1e99):
t = time.time()
while self.i_epoch < n_epoch:
while time.time() - t < 10 and self.i_epoch < n_epoch:
# Target policy as player 1
self.TrainOneRound(
self.target_policy.move_dict[self.start_num])
self.i_epoch += 1
# Target policy as player 2
self.TrainOneRound(self.start_num)
t = time.time()
pickle.dump((self.target_policy, self.i_epoch),
open(self.path, "wb"))
print("Trained %i epochs so far." % self.i_epoch)
def TrainOneRound(self, afterstate_num, alpha=0.1):
""" TD(0) following Sutton and Barto 6.1
Input:
afterstate: the afterstate of target_policy to start trainng with
Note that the opponent mamkes a move first, then the target policy.
"""
afterstate = State(from_base10=afterstate_num)
while not afterstate.is_terminal():
beforestate_num = self.opponent_policy.move(afterstate.get_num())
beforestate = State(from_base10=beforestate_num)
if beforestate.is_terminal():
r = beforestate.get_reward()
self.target_policy.v_dict[afterstate.get_num()] += alpha * (
r - self.target_policy.v_dict[afterstate.get_num()])
break
else:
s_prime_num = self.target_policy.move(beforestate_num)
s_prime = State(from_base10=s_prime_num)
r = s_prime.get_reward()
self.target_policy.v_dict[afterstate.get_num()] += alpha * (
r + self.target_policy.v_dict[s_prime_num] -
self.target_policy.v_dict[afterstate.get_num()])
afterstate = s_prime
def test_get_trajectory():
trainer = Train(path='foo', read_first=False)
trainer.epsilon = 0
trajectory = trainer.GetOneTrajectory(TabularPolicy(), TabularPolicy())
num1 = State(board=[[0, 0, 0], [0, 0, 0], [0, 0, 0]]).get_num()
num2 = State(board=[[1, 0, 0], [0, 0, 0], [0, 0, 0]], turn=2).get_num()
num3 = State(board=[[1, 2, 0], [0, 0, 0], [0, 0, 0]]).get_num()
num4 = State(board=[[1, 2, 1], [0, 0, 0], [0, 0, 0]], turn=2).get_num()
num5 = State(board=[[1, 2, 1], [2, 0, 0], [0, 0, 0]]).get_num()
num6 = State(board=[[1, 2, 1], [2, 1, 0], [0, 0, 0]], turn=2).get_num()
num7 = State(board=[[1, 2, 1], [2, 1, 2], [0, 0, 0]]).get_num()
num8 = State(board=[[1, 2, 1], [2, 1, 2], [1, 0, 0]], turn=2).get_num()
assert trajectory == [num1, num2, num3, num4, num5, num6, num7, num8]
def __init__(self, path, read_first=False):
"""
Input:
path: the path to save the policy
read_first: if true, read from the path first
"""
if read_first:
self.policy_1, self.i_epoch = pickle.load(open(path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.policy_1 = TabularPolicy()
self.i_epoch = 0
self.path = path
self.policy_stable = True
class Train:
def __init__(self, path, read_first=False, epsilon=.9):
"""
Input:
path: the path to save the policy
read_first: if true, read from the path first
"""
if read_first:
self.policy_1, self.i_epoch, self.returns = pickle.load(
open(path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.policy_1 = TabularPolicy(epsilon=epsilon)
self.i_epoch = 0
self.returns = dict()
self.path = path
self.policy_stable = True
self.epsilon = epsilon
self.policy_1.epsilon = epsilon
def OnPolicyMCControl(self):
""" On-policy MC control following Sutton Barto 5.4
"""
t = time.time()
while True:
while time.time() - t < 10:
num = State().get_num()
history = [num]
while not State(from_base10=num).is_terminal():
num = self.policy_1.move(num)
history.append(num)
# g is a constant for our case
g = State(from_base10=num).get_reward()
for i, num in enumerate(history):
if num in self.returns:
self.returns[num].append(g)
else:
self.returns[num] = [g]
self.policy_1.v_dict[num] = np.average(self.returns[num])
if self.policy_1.be_greedy(history):
self.policy_stable = False
self.i_epoch += 1
t = time.time()
pickle.dump((self.policy_1, self.i_epoch, self.returns),
open(self.path, "wb"))
print("Trained %i epochs so far." % self.i_epoch)
def MCPrediction(self, n_epoch):
""" MC prediction following Sutton Barto 5.1
Against rush opponent
Input:
n_epoch: the number of episodes to be trained
"""
self.policy_2 = TabularPolicy()
returns = dict()
for num in range(int('1' + '0' * 9, 3), int('2' * 10, 3) + 1):
returns[num] = []
for _ in range(n_epoch):
# generate an episode following policy_1
s = State().get_num()
history = [s]
while not State(from_base10=s).is_terminal():
s = self.policy_1.move_dict[s]
history.append(s)
if State(from_base10=s).is_terminal():
break
s = self.policy_2.move_dict[s]
history.append(s)
# in our special case, g is a constant
g = State(from_base10=s).get_reward()
for i, s in enumerate(history):
returns[s].append(g)
if i % 2 == 0:
self.policy_1.v_dict[s] = np.average(returns[s])
else:
self.policy_2.v_dict[s] = np.average(returns[s])
for num in range(int('2' + '0' * 9, 3), int('2' * 10, 3) + 1):
self.policy_1.v_dict[num] = self.policy_2.v_dict[num]
self.i_epoch += 1
pickle.dump((self.policy_1, self.i_epoch), open(self.path, "wb"))
print('MC prediction finished.')
def __init__(self, path, read_first=False):
"""
Input:
path: the path to save the policy
read_first: if true, read from the path first
"""
if read_first:
self.target_policy, self.i_epoch = pickle.load(open(path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.target_policy = TabularPolicy()
self.i_epoch = 0
self.opponent_policy = TabularPolicy()
self.path = path
# num for the state with an empty board and with player 1 to make a move.
self.start_num = int('1' + '0' * 9, 3)
def __init__(self, write_path, read_path=None, self_play=False):
"""
Input:
n_game: number of games to train for
read_path, write_path: paths for reading or saving the model
"""
self.read_path = read_path
self.write_path = write_path
if read_path:
self.policy_1, self.i_epoch = pickle.load(
open(self.read_path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.policy_1 = TabularPolicy()
self.i_epoch = 0
print('Training new policy.')
self.read_path = self.write_path # for later iterative training
if self_play:
self.policy_2 = self.policy_1
else:
self.policy_2 = TabularPolicy()
self.policy_ever_changed = True # Set to true to state iterative training
class TrainOneRound:
def __init__(self, path, read_first=False):
"""
Input:
path: the path to save the policy
read_first: if true, read from the path first
"""
if read_first:
self.policy_1, self.i_epoch = pickle.load(open(path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.policy_1 = TabularPolicy()
self.i_epoch = 0
self.path = path
self.policy_stable = True
def MCES(self):
""" MC exploring start following Sutton Barto 5.3
Against rush opponent
"""
t = time.time()
# No need to use a list of returns, since the game is deterministic
for s in range(int('1' + '0' * 9, 3), int('2' * 10, 3) + 1):
# print(self.policy_1.v_dict[47042])
history = [s]
while not State(from_base10=s).is_terminal():
s = self.policy_1.move_dict[s]
history.append(s)
if State(from_base10=s).is_terminal():
break
s = self.policy_1.move_dict[s]
history.append(s)
g = State(from_base10=s).get_reward()
for i, s in enumerate(history):
self.policy_1.v_dict[s] = g
if self.policy_1.be_greedy(history):
self.policy_stable = False
self.i_epoch += 1
if time.time() - t > 10:
t = time.time()
pickle.dump((self.policy_1, self.i_epoch),
open(self.path, "wb"))
print("Trained %i epochs so far." % self.i_epoch)
pickle.dump((self.policy_1, self.i_epoch), open(self.path, "wb"))
print('MC exploring start finished.')
class TrainOneRound:
def __init__(self, write_path, read_path=None, self_play=False):
"""
Input:
n_game: number of games to train for
read_path, write_path: paths for reading or saving the model
"""
self.read_path = read_path
self.write_path = write_path
if read_path:
self.policy_1, self.i_epoch = pickle.load(
open(self.read_path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.policy_1 = TabularPolicy()
self.i_epoch = 0
print('Training new policy.')
self.read_path = self.write_path # for later iterative training
if self_play:
self.policy_2 = self.policy_1
else:
self.policy_2 = TabularPolicy()
self.policy_ever_changed = True # Set to true to state iterative training
def ValueIteration(self, theta=0.01):
t = time.time()
while True:
delta = 0
for num in range(int('1' + '0' * 9, 3), int('2' * 10, 3) + 1):
v = self.policy_1.v_dict[num]
state = State(from_base10=num)
if state.is_terminal():
self.policy_1.v_dict[num] = state.get_reward()
else:
opponent_afterstate = State(
from_base10=self.policy_2.move_dict[num])
if opponent_afterstate.is_terminal():
self.policy_1.v_dict[
num] = opponent_afterstate.get_reward()
else:
s_prime_choices = opponent_afterstate.legal_afterstates(
)
if state.turn == 2:
vi_update = max([
self.policy_1.v_dict[x]
for x in s_prime_choices
])
else:
vi_update = min([
self.policy_1.v_dict[x]
for x in s_prime_choices
])
self.policy_1.v_dict[num] = vi_update
delta = max(delta, np.abs(v - self.policy_1.v_dict[num]))
self.i_epoch += 1
if delta < theta:
print('Value function has converged!')
print("Trained %i epochs so far." % self.i_epoch)
self.policy_ever_changed = self.policy_1.be_greedy()
pickle.dump((self.policy_1, self.i_epoch),
open(self.write_path, "wb"))
break
if time.time() - t > 10:
t = time.time()
print("Trained %i epochs so far." % self.i_epoch)
self.policy_ever_changed = self.policy_1.be_greedy()
pickle.dump((self.policy_1, self.i_epoch),
open(self.write_path, "wb"))
@author: daugh
"""
from ttt_play import State
from ttt_policies import TabularPolicy
import os
import pickle
policy, i_epoch = pickle.load(
open(os.path.dirname(os.getcwd()) + '/policy_evaluation.pkl', 'rb'))
print('This value function has been trained for %i epochs.' % i_epoch)
theta = 0.01
print('Accuracy %f' % theta)
opponent_policy = TabularPolicy(epsilon=1)
results = []
for i in range(1000):
state = State()
while True:
state = State(from_base10=policy.move_dict[state.get_num()])
if state.is_terminal():
break
else:
state = State(from_base10=opponent_policy.move(state.get_num()))
if state.is_terminal():
break
results.append(state.get_reward())
print("Average reward %f over 1000 games as player X against random policy." %
(sum(results) / 1000.))
class Train:
def __init__(self, path, read_first=False):
"""
Input:
path: the path to save the policy
read_first: if true, read from the path first
"""
if read_first:
self.target_policy, self.i_epoch, self.c = pickle.load(
open(path, 'rb'))
print('Policy read from file. Trained for %i epochs.' % self.i_epoch)
else:
self.target_policy = TabularPolicy()
self.i_epoch = 0
self.c = {}
self.behavior_policy = TabularPolicy(epsilon=1)
self.path = path
self.policy_stable = True
def TrainContinuously(self, n_epoch=1e99):
t = time.time()
while self.i_epoch < n_epoch:
while time.time() - t < 10 and self.i_epoch < n_epoch:
self.TrainOneRound()
self.i_epoch += 1
t = time.time()
pickle.dump((self.target_policy, self.i_epoch, self.c),
open(self.path, "wb"))
print("Trained %i epochs so far." % self.i_epoch)
def TrainOneRound(self):
""" Off-policy MC prediction following Sutton Barto 5.6
"""
# behavior policy playing player 1
trajectory = self.GetOneTrajectory(
self.behavior_policy, self.target_policy)
self.OffPolicyMCControl(trajectory, 1)
# behavior policy playing player 2
trajectory = self.GetOneTrajectory(
self.target_policy, self.behavior_policy)
self.OffPolicyMCControl(trajectory, 2)
def GetOneTrajectory(self, policy_1, policy_2):
"""
Returns: list of state nums of a trajectory
"""
num = State().get_num()
trajectory = [num]
while not State(from_base10=num).is_terminal():
num = policy_1.move(num)
trajectory.append(num)
if not State(from_base10=num).is_terminal():
num = policy_2.move(num)
trajectory.append(num)
else:
break
return trajectory
def OffPolicyMCControl(self, trajectory, role_behavior_policy):
""" Incremental implementation of off-policy MC prediction
Input:
trajectory
role_behavior_policy: 1 or 2, denoting which player the behavior policy acted as in this trajectory
"""
# g is a constant for our case
g = State(from_base10=trajectory[-1]).get_reward()
w = 1.
i = len(trajectory) - 1
for i, state in reversed(list(enumerate(trajectory))):
if i == len(trajectory) - 1:
# ignore the very last state, which is not a beforestate
continue
if (i % 2 + 1) != role_behavior_policy:
# i denotes the number of pieces on the board. i%2+1 is 1 if
# this is player 1's before state, and is 2 if this is player
# 2's before state.
continue
afterstate = trajectory[i+1]
if afterstate in self.c:
self.c[afterstate] += w
else:
self.c[afterstate] = w
self.target_policy.v_dict[afterstate] += w / \
self.c[afterstate] * \
(g - self.target_policy.v_dict[afterstate])
self.target_policy.be_greedy([state])
if self.target_policy.move_dict[trajectory[i]] != afterstate:
break
else:
w = w * \
len(State(from_base10=trajectory[i]).legal_afterstates())
class Train:
def __init__(self, path, read_first=False):
"""
Input:
path: the path to save the policy
read_first: if true, read from the path first
"""
if read_first:
self.target_policy, self.i_epoch = pickle.load(open(path, 'rb'))
print('Policy read from file. Trained for %i epochs.' %
self.i_epoch)
else:
self.target_policy = TabularPolicy()
self.i_epoch = 0
self.random_policy = TabularPolicy(epsilon=1)
self.path = path
# num for the state with an empty board and with player 1 to make a move.
self.start_num = int('1' + '0' * 9, 3)
def TrainContinuously(self, n_epoch=1e99):
t = time.time()
while self.i_epoch < n_epoch:
while time.time() - t < 10 and self.i_epoch < n_epoch:
# Target policy as player 1
self.TrainOneRound(self.random_policy.move(self.start_num))
self.i_epoch += 1
# Target policy as player 2
self.TrainOneRound(self.start_num)
t = time.time()
self.target_policy.be_greedy()
pickle.dump((self.target_policy, self.i_epoch),
open(self.path, "wb"))
print("Trained %i epochs so far." % self.i_epoch)
def TrainOneRound(self, afterstate_num, alpha=.1):
""" Q learning following Sutton and Barto 6.5
Input:
afterstate: the afterstate of target_policy to start trainng with
Note that the opponent makes a move first, then the target policy.
"""
afterstate = State(from_base10=afterstate_num)
while not afterstate.is_terminal():
beforestate_num = self.random_policy.move(
afterstate.get_num()) # opponent makes a move
beforestate = State(from_base10=beforestate_num)
if beforestate.is_terminal():
r = beforestate.get_reward()
self.target_policy.v_dict[afterstate.get_num()] += alpha * (
r - self.target_policy.v_dict[afterstate.get_num()])
break
else:
s_primes = beforestate.legal_afterstates()
candidates = []
for s_prime in s_primes:
r = State(from_base10=s_prime).get_reward()
q = self.target_policy.v_dict[s_prime]
candidates.append(r + q)
if beforestate.turn == 1:
self.target_policy.v_dict[
afterstate.get_num()] += alpha * (
max(candidates) -
self.target_policy.v_dict[afterstate.get_num()])
else:
self.target_policy.v_dict[
afterstate.get_num()] += alpha * (
min(candidates) -
self.target_policy.v_dict[afterstate.get_num()])
afterstate_num = self.random_policy.move(beforestate_num)
afterstate = State(from_base10=afterstate_num)