def generate_vi(mdp, c, lr=0.1):
init_pi = utils.random_policy(mdp.S,mdp.A)
init_v = utils.value_functional(mdp.P, mdp.r, init_pi, mdp.discount)
vs = np.stack(utils.solve(ss.value_iteration(mdp, lr), init_v))[:,:,0]
n = vs.shape[0]
plt.scatter(vs[0, 0], vs[0, 1], c=c, s=30, label='{}'.format(n))
plt.scatter(vs[1:-1, 0], vs[1:-1, 1], c=range(n-2), cmap='viridis', s=10)
plt.scatter(vs[-1, 0], vs[-1, 1], c='m', marker='x')
def generate_pg(mdp, c, lr=0.01):
init_pi = utils.random_policy(mdp.S,mdp.A)
init_logit = np.log(init_pi)
logits = utils.solve(ss.policy_gradient_iteration_logits(mdp, lr), init_logit)
vs = np.stack([utils.value_functional(mdp.P, mdp.r, utils.softmax(logit), mdp.discount) for logit in logits])[:,:,0]
n = vs.shape[0]
plt.scatter(vs[0, 0], vs[0, 1], c=c, s=30, label='{}'.format(n))
plt.scatter(vs[1:-1, 0], vs[1:-1, 1], c=range(n-2), cmap='viridis', s=10)
plt.scatter(vs[-1, 0], vs[-1, 1], c='m', marker='x')
def generate_model_cs():
"""
Compare using all deterministic policies versus fewer mixed policies.
Starts to get interesting in higher dims?
"""
n_states = 32
n_actions = 2
lr = 0.01
k = 64
mdp = utils.build_random_mdp(n_states, n_actions, 0.5)
init = rnd.standard_normal((mdp.S * mdp.S * mdp.A + mdp.S * mdp.A))
pi_star = utils.solve(policy_iteration(mdp),
utils.softmax(rnd.standard_normal(
(mdp.S, mdp.A))))[-1]
print('pi_star\n', pi_star)
# adversarial pis
# apis = utils.get_deterministic_policies(mdp.S, mdp.A)
apis = np.stack([utils.random_det_policy(mdp.S, mdp.A) for _ in range(k)])
update_fn = model_iteration(mdp, lr, apis)
params = utils.solve(update_fn, init)
p_logits, r = parse_model_params(mdp.S, mdp.A, params[-1])
error = np.mean(
(utils.value_functional(mdp.P, mdp.r, pi_star, mdp.discount) -
utils.value_functional(utils.softmax(p_logits), r, pi_star,
mdp.discount))**2)
print('\n', error)
new_mdp = utils.MDP(mdp.S, mdp.A, utils.softmax(p_logits), r, mdp.discount,
mdp.d0)
pi_star = utils.solve(policy_iteration(new_mdp),
utils.softmax(rnd.standard_normal(
(mdp.S, mdp.A))))[-1]
print(pi_star)
apis = np.stack([utils.random_policy(mdp.S, mdp.A) for _ in range(k)])
update_fn = model_iteration(mdp, lr, apis)
params = utils.solve(update_fn, init)
p_logits, r = parse_model_params(mdp.S, mdp.A, params[-1])
error = np.mean(
(utils.value_functional(mdp.P, mdp.r, pi_star, mdp.discount) -
utils.value_functional(utils.softmax(p_logits), r, pi_star,
mdp.discount))**2)
print('\n', error)
new_mdp = utils.MDP(mdp.S, mdp.A, utils.softmax(p_logits), r, mdp.discount,
mdp.d0)
pi_star = utils.solve(policy_iteration(new_mdp),
utils.softmax(rnd.standard_normal(
(mdp.S, mdp.A))))[-1]
print(pi_star)
def k_step_option_similarity():
n_states, n_actions = 6, 2
mdp = utils.build_random_mdp(n_states, n_actions, 0.5)
pi = utils.random_policy(n_states, n_actions)
P = multi_step_transition_fn(mdp.P, pi, 3)
# P[:,-1] = P[:,-2]
# s(o1, o2) = sum_s' P(s' | s1) * log( P(s' | s2) / P(s' | s1))
kl = -np.sum(P[:, :, None] * np.log(P[:, None, :] / P[:, :, None]), axis=0)
print(kl)
def generate_pi(mdp, c):
init_pi = utils.random_policy(mdp.S,mdp.A)
pis = utils.solve(ss.policy_iteration(mdp), init_pi)
vs = np.stack([utils.value_functional(mdp.P, mdp.r, pi, mdp.discount) for pi in pis])[:,:,0]
n = vs.shape[0]
plt.scatter(vs[0, 0], vs[0, 1], c=c, s=30, label='{}'.format(n-2))
plt.scatter(vs[1:-1, 0], vs[1:-1, 1], c=range(n-2), cmap='viridis', s=10)
plt.scatter(vs[-1, 0], vs[-1, 1], c='m', marker='x')
for i in range(len(vs)-2):
dv = 0.1*(vs[i+1, :] - vs[i, :])
plt.arrow(vs[i, 0], vs[i, 1], dv[0], dv[1], color=c, alpha=0.5, width=0.005)
def emp_est_snr_graph():
n_states, n_actions = 12, 3
mdp = utils.build_random_mdp(n_states, n_actions, 0.5)
pis = [utils.random_policy(n_states, n_actions) for _ in range(100)]
vs = []
hs = []
for i, pi in enumerate(pis):
print('\r{}'.format(i), end='', flush=True)
# try:
vs.append(est_var_R(mdp, pi))
hs.append(utils.entropy(pi))
# except ValueError as err:
# print(err)
plt.scatter(hs, vs)
plt.show()
def state_action_vis():
# want to pick policies that maximise exploration.
# but. how to solve for this analytically?! not sure this is going to work...
# unless? is there a way to analytically set pi = 1/visitation?!
# if we iterate. estimate visitation under pi, set pi = 1/visitaiton.
# does it converge? where does it converge?
# it shouldnt converge?!?
mdp = utils.build_random_mdp(12, 2, 0.5)
pi = utils.random_policy(mdp.S, mdp.A)
v_sa_sa = state_action_visitation_distribution(mdp, pi)
# sum over initial conditions to get discounted state-action visitation probability
d0_sa = np.reshape(np.einsum('jk,jl->jk', pi, mdp.d0), (mdp.S * mdp.A, ))
ps = np.einsum('ik,k->i', v_sa_sa, d0_sa)
plt.imshow(v_sa_sa)
plt.show()
def generate_snr_map():
n_states, n_actions = 2, 3
mdp = utils.build_random_mdp(n_states, n_actions, 0.5)
# pis = utils.gen_grid_policies(11)
pis = [utils.random_policy(n_states, n_actions) for _ in range(512)]
Vs = utils.polytope(mdp.P, mdp.r, mdp.discount, pis)
mags = [grad_mag(mdp.P, mdp.r, pi, mdp.discount) for pi in pis]
uncert = [variance(mdp.P, mdp.r, pi, mdp.discount) for pi in pis]
snr = [s / n for s, n in zip(mags, uncert)]
plt.subplot(3, 1, 1)
plt.title('Magnitude')
plt.scatter(Vs[:, 0], Vs[:, 1], c=mags)
plt.subplot(3, 1, 2)
plt.title('Variance')
plt.scatter(Vs[:, 0], Vs[:, 1], c=uncert)
plt.subplot(3, 1, 3)
plt.title('SNR')
plt.scatter(Vs[:, 0], Vs[:, 1], c=snr)
plt.show()
# left
P[:, :, 3] = np.array([
[1, 1, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 1, 1],
[0, 0, 0, 0, 0, 0],
])
# rewards. 6 x 4
r = np.array([
[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0],
[1, 0, 0, 0], # rewarded for going up at the finish.
[1, 0, 0, 0],
])
# initial distribution
d0 = np.array([[0.5, 0.5, 0, 0, 0, 0]])
pi = np.array(utils.random_policy(6, 4))
pi[[0, 2, 4]] = pi[[1, 3, 5]]
V = utils.value_functional(P, r, pi, 0.5)
Q_t = utils.bellman_operator(P, r, V, 0.5)
# print(np.sum(P, axis=-1))
print(Q_t)
Can we do TD with values from different MDPs?
"""
V_true = vmap(
lambda pi: utils.value_functional(mdp.P, mdp.r, pi, mdp.discount))
V_guess = vmap(lambda params, pi: utils.value_functional(
*mdp_sampler(params), pi, mdp.discount),
in_axes=(None, None, 0))
dLdp = grad(lambda params: mse(V_true(pis), V_guess(params, pis)))
@jit
def update_fn(params, Q):
m = symmetric_sampler(params)
Q_ = utils.bellman_optimality_operator(m.P, m.r, Q, m.discount)
params_tp1 -= lr * dLdp(
params
) # done based on observations... could use model iteration!?
Q_tp1 = Q + lr * (Q_ - Q)
return Q_tp1, params_tp1
return update_fn
if __name__ == "__main__":
# np.random.seed(0)
n_states, n_actions = 16, 2
mdp = utils.build_random_mdp(n_states, n_actions, 0.5)
pis = [utils.random_policy(n_states, n_actions) for _ in range(1000)]
# pis = utils.get_deterministic_policies(n_states, n_actions)
onoffpolicy_abstraction(mdp, pis)
