def qlearning_grid(lr, eps, epoch=100, searchbound=400):
grid = Grid()
grid.pi_params = np.zeros((23, 4))
grid.softmax()
actions = grid.action
estimated_rewards = np.zeros(epoch)
q = np.zeros((23, 4))
for x in range(epoch):
s = grid.d_zero()
while s != [5, 5]:
# choose new_a from new_s using policy derived from q
pi_temp = pe.softmax(q[grid.get_index(s)], actions, eps(x))
a = np.random.choice(actions, 1, p=pi_temp)[0]
# print(q)
# Take action a and observe r and s′;
new_s, r = grid.P_and_R(s, a)
q[grid.get_index(s), actions.index(a)] += lr * (
r + grid.gamma * np.max(q[grid.get_index(new_s)]) -
q[grid.get_index(s), actions.index(a)])
s = new_s
# using q function to estimate the reward and add it to estimated_reward
# print('episode: ', x, ', q function: ', q)
grid.pi_params = pe.softmax(q, actions, eps(x))
grid_epi = GridEpisode(grid, step_bound=searchbound)
# print('episode: ', x, ', pi: ', grid.pi_params)
estimated_rewards[x] = grid_epi.run_all_steps()
print('episode: ', x, ', reward: ', estimated_rewards[x], 'epsilon: ',
eps(x))
# decay *= decay_rate
return estimated_rewards
def sarsa_cartpole(lr, baseparams, epoch=100, eps=1e-2, base='fourier'):
cartpole = CartPole()
estimated_rewards = np.zeros(epoch)
actions = cartpole.actions
w = None
if base == 'fourier':
order = baseparams['order']
s = cartpole.d_zero()
w = np.zeros((1, len(actions) * (order + 1)**len(s)))
elif base == 'tile':
num_tilings, tiles_per_tiling = baseparams['num_tilings'], baseparams[
'tiles_per_tiling']
s = cartpole.d_zero()
w = np.zeros((1, len(actions) * num_tilings))
for x in range(epoch):
s = cartpole.d_zero()
# choose a from s using a policy derived from q (e.g., ε-greedy or softmax);
first_q = pe.softmax(pe.qw(w, s, actions, base, baseparams), actions,
eps)
# pi_s = pe.epsilon_greedy(pe.qw(w, s, order, actions, base), actions, eps)
a = np.random.choice(actions, 1, p=first_q)[0]
count = 0
while np.abs(s[0]) < cartpole.edge and np.abs(
s[1]) < cartpole.fail_angle and count < 1010:
# Take action a and observe r and s′;
new_s, r = cartpole.P_and_R(s, a)
# Choose a′ from s′ using a policy derived from q;
pi_temp = pe.softmax(pe.qw(w, new_s, actions, base, baseparams),
actions, eps)
new_a = np.random.choice(actions, 1, p=pi_temp)[0]
# w += lr * (r + pe.qw_fourier_ele(w, new_s, new_a, order, actions) -
# pe.qw_fourier_ele(w, s, a, order, actions)) * pe.dqwdw_fourier(s, a, order, actions)
new_q = pe.qw_ele(w, new_s, new_a, actions, base, baseparams)[0]
q, dqdw = pe.qw_ele(w, s, a, actions, base, baseparams)
w += lr * (r + new_q - q) * dqdw
s = new_s
a = new_a
count += 1
epi = CartPoleEpisode(cartpole)
estimated_rewards[x] = epi.run_with_w_softmax(w, eps, base, baseparams)
print('episode: ', x, ', reward: ', estimated_rewards[x])
# print('episode: ', x, ', w: ', w)
return estimated_rewards
def reinforce_grid(lr, eps, epoch=100, searchbound=400):
estimated_rewards = np.zeros(epoch)
# theta is a representation of policy
theta = np.zeros((23, 4))
grid = Grid()
actions = grid.action
# print(epoch)
# for each episode:
for x in range(epoch):
# s ∼ d0
s = grid.d_zero()
count = 0
hist_s = []
hist_a = []
hist_r = []
grid.pi_params = estimation.softmax(theta, eps(x))
# for each time step, until s is the terminal absorbing state do
while s != [5, 5] and count < 1000:
hist_s.append(s)
a = grid.pi(s)
hist_a.append(a)
new_s, r = grid.P_and_R(s, a)
hist_r.append(r)
s = new_s
count += 1
# delta_j = 0
decay = 1
for i in range(len(hist_s)):
g = 0
gd = 1
for j in range(i, len(hist_s)):
g += gd * hist_r[j]
gd *= grid.gamma
theta[grid.get_index(hist_s[i]),
actions.index(hist_a[i])] += lr * decay * g
decay *= grid.gamma
grid.pi_params = estimation.softmax(theta, eps(x))
# grid.softmax()
grid_epi = GridEpisode(grid, step_bound=searchbound)
# print('episode: ', x, ', pi: ', grid.pi_params)
estimated_rewards[x] = grid_epi.run_all_steps()
if x == epoch - 1:
print('episode: ', x, ', reward: ', estimated_rewards[x])
# decay *= decay_rate
return estimated_rewards
def run_with_w_softmax(self, w, eps, base, baseparams):
reward = 0
s = self.mountaincar.d_zero()
count = 0
while self.active == 1:
q = fa.qw(w, s, self.mountaincar.actions, base, baseparams)
# print(q)
pi = estimation.softmax(q, eps)
# pi = estimation.epsilon_greedy(q, self.mountaincar.actions, eps)
# print('test')
# print(pi)
a = np.random.choice(self.mountaincar.actions, 1, p=pi)[0]
# print(pi, s, a)
# print(a)
# print(s[0])
s, r = self.mountaincar.P_and_R(s, a)
count += 1
if s[0] == self.mountaincar.right_bound:
self.active = 0
else:
reward += r
if count >= 1e3:
self.active = 0
return reward
def actor_critic_grid(lr, eps, epoch=100, searchbound=400):
estimated_rewards = np.zeros(epoch)
# Initialize tabular-v arbitrarily
v = np.zeros(23)
# theta is a representation of policy
theta = np.zeros((23, 4))
grid = Grid()
actions = grid.action
# for each episode:
for x in range(epoch):
# s ∼ d0
s = grid.d_zero()
count = 0
# for each time step, until s is the terminal absorbing state do
while s != [5, 5] and count < 1000:
# a ∼ π(s, ·);
grid.pi_params = estimation.softmax(theta, eps(x))
a = grid.pi(s)
# Take action a and observe r and s′;
new_s, r = grid.P_and_R(s, a)
# Critic update using TD(λ)
# e ← γλe + ∂qw(s,a)/∂qw;
delta = r + grid.gamma * v[grid.get_index(new_s)] - v[
grid.get_index(s)]
# w←w+αδev;
v[grid.get_index(s)] += lr * delta
theta[grid.get_index(s), actions.index(a)] += lr * delta
# print(theta)
s = new_s
count += 1
# using q function to estimate the reward and add it to estimated_reward
# print('episode: ', x, ', q function: ', q)
grid.pi_params = estimation.softmax(theta, eps(x))
# grid.softmax()
grid_epi = GridEpisode(grid, step_bound=searchbound)
# print('episode: ', x, ', pi: ', grid.pi_params)
estimated_rewards[x] = grid_epi.run_all_steps()
if x == 99:
print('episode: ', x, ', reward: ', estimated_rewards[x])
# decay *= decay_rate
return estimated_rewards
def run_with_w_softmax(self, w, eps, base, baseparams):
reward = 0
s = self.cartpole.d_zero()
while self.active == 1:
q = pe.qw(w, s, self.cartpole.actions, base, baseparams)
pi = pe.softmax(q, self.cartpole.actions, eps)
a = np.random.choice(self.cartpole.actions, 1, p=pi)[0]
s, r = self.cartpole.P_and_R(s, a)
reward += 1
self.step_count += 1
if self.step_count > self.maxturn:
self.active = 0
if np.abs(s[0]) > self.cartpole.edge:
self.active = 0
if np.abs(s[1]) > self.cartpole.fail_angle:
self.active = 0
return reward - 1
def reinforce_mc(alpha, beta, l, baseparams, eps, epoch=100, base='fourier'):
mc = MountainCar()
estimated_rewards = np.zeros(epoch)
actions = mc.actions
theta = None
order = 0
if base == 'fourier':
order = baseparams['order']
s = mc.d_zero()
theta = np.zeros((1, len(actions) * (order + 1)**len(s)))
w = np.zeros((1, (order + 1)**len(s)))
# theta = np.zeros((len(s), 3))
for x in range(epoch):
s = mc.d_zero()
e = np.zeros(w.shape)
hist_s = []
hist_a = []
hist_r = []
hist_pi = []
count = 0
dj = np.zeros(theta.shape)
# for each time step, until s is the terminal absorbing state do
while s[0] < mc.right_bound and count < 1000:
pi_temp = estimation.softmax(
fa.qw(theta, s, actions, base, baseparams), eps(x))
a = np.random.choice(actions, 1, p=pi_temp)[0]
new_s, r = mc.P_and_R(s, a)
hist_a.append(a)
hist_s.append(s)
hist_r.append(r)
hist_pi.append(pi_temp)
s = new_s
count += 1
for i in range(len(hist_a)):
g = 0
for j in range(i, len(hist_s)):
g += hist_r[j]
v, dv = fa.vw(w, hist_s[i], base, baseparams)
dj += (g - v) * dsoftmax(hist_s[i], hist_a[i], order, actions,
hist_pi[i])
e = l * e + dv
if i == len(hist_s) - 1:
delta = hist_r[i] + 0 - \
fa.vw(w, hist_s[i], base, baseparams)[0]
else:
delta = hist_r[i] + fa.vw(w, hist_s[i + 1], base, baseparams)[0] - \
fa.vw(w, hist_s[i], base, baseparams)[0]
w += alpha * delta * e
theta += beta * dj
epi = MountainCarEpisode(mc)
# print(theta)
estimated_rewards[x] = epi.run_with_w_softmax(theta, eps(x), base,
baseparams)
print('episode: ', x, ', reward: ', estimated_rewards[x])
return estimated_rewards
def actor_critic_mc(lr, l, baseparams, eps, epoch=100, base='fourier'):
mc = MountainCar()
estimated_rewards = np.zeros(epoch)
actions = mc.actions
w = None
theta = None
order = 0
if base == 'fourier':
order = baseparams['order']
s = mc.d_zero()
w = np.zeros((1, (order + 1)**len(s)))
theta = np.zeros((1, len(actions) * (order + 1)**len(s)))
# theta = np.zeros((len(s), 3))
for x in range(epoch):
s = mc.d_zero()
# ev ← 0
e = np.zeros(w.shape)
# et ← 0
# et = np.zeros(theta.shape)
count = 0
# for each time step, until s is the terminal absorbing state do
while s[0] < mc.right_bound and count < 1000:
pi_temp = estimation.softmax(
fa.qw(theta, s, actions, base, baseparams), eps(x))
a = np.random.choice(actions, 1, p=pi_temp)[0]
# print(a)
# print(pi_temp)
# dydtheta_list = []
# for na in actions:
# dydtheta_list.append(fa.qw_ele(theta, s, na, actions, base, baseparams)[1])
#
# dtheta = estimation.dsoftmax(fa.qw(theta, s, actions, base, baseparams), dydtheta_list, actions.index(
# a), eps(x))
dtheta = np.zeros((1, len(actions) * (order + 1)**len(s)))
for idx in range(len(actions)):
phi = fa.fourier_phi_mc(s, order).T
if actions[idx] == a:
# print('target')
dtheta[:, idx * phi.shape[1]:(idx + 1) *
phi.shape[1]] = (1 - pi_temp[idx]) * phi
else:
dtheta[:, idx * phi.shape[1]:(idx + 1) *
phi.shape[1]] = -pi_temp[idx] * phi
# Take action a and observe r and s′;
new_s, r = mc.P_and_R(s, a)
# Critic update using TD(λ)
# ev←γλev+∂vw(s);
v, dv = fa.vw(w, s, base, baseparams)
if new_s[0] > mc.right_bound:
new_v = 0
else:
new_v = fa.vw(w, new_s, base, baseparams)[0]
e = l * mc.gamma * e
e += dv
# δ ← r + γvw(s′,a′) − vw(s,a);
delta = r + mc.gamma * new_v - v
# w←w+αδev;
w += lr * delta * e
# Actor update
# θ + αγ^tδ ∂ ln(π(s,a,θ))
theta += lr * delta * dtheta
s = new_s
count += 1
epi = MountainCarEpisode(mc)
# print(theta)
estimated_rewards[x] = epi.run_with_w_softmax(theta, eps(x), base,
baseparams)
print('episode: ', x, ', reward: ', estimated_rewards[x])
return estimated_rewards