def modified_policy_iteration(mdp: MDP,
gamma: float,
epsilon: float,
k: int = 5) -> Tuple[Dict, Dict]:
random_a = random.sample(mdp.A, 1)[0]
pi = {s: random_a for s in mdp.S}
V = {s: 0. for s in mdp.S}
while True:
for i in range(k):
for s in mdp.S:
V[s] = mdp.R(s, pi[s]) + gamma * sum([
mdp.P(s_prime, s, pi[s]) * V[s_prime] for s_prime in mdp.S
])
delta = 0.
for s in mdp.S:
V_old = V[s]
V_new = {
a: mdp.R(s, a) + gamma *
sum([mdp.P(s_prime, s, a) * V[s_prime] for s_prime in mdp.S])
for a in mdp.A
}
pi[s] = max(V_new, key=V_new.get)
V[s] = max(V_new.values())
delta = max(delta, abs(V[s] - V_old))
if delta <= epsilon:
break
return pi, V
def async_value_iteration(mdp: MDP,
gamma: float,
num_iterations: int = 1000) -> Tuple[Dict, Dict]:
Q = {(s, a): 0. for a in mdp.A for s in mdp.S}
for i in range(num_iterations):
s = random.sample(mdp.S, 1)[0]
a = random.sample(mdp.A, 1)[0]
Q[(s, a)] = mdp.R(s, a) + gamma * sum([
mdp.P(s_prime, s, a) *
max([Q[(s_prime, a_prime)] for a_prime in mdp.A])
for s_prime in mdp.S
])
pi = {}
for s in mdp.S:
values = {a: Q[(s, a)] for a in mdp.A}
pi[s] = max(values, key=values.get)
return pi, Q
def value_iteration(mdp: MDP, gamma: float,
epsilon: float) -> Tuple[Dict, Dict]:
V = {s: (None, 0.) for s in mdp.S}
while True:
delta = 0.
for s in mdp.S:
V_old = V[s][1]
V_new = {
a: mdp.R(s, a) + gamma * sum([
mdp.P(s_prime, s, a) * V[s_prime][1] for s_prime in mdp.S
])
for a in mdp.A
}
V[s] = (max(V_new, key=V_new.get), max(V_new.values()))
delta = max(delta, abs(V[s][1] - V_old))
if delta <= epsilon:
break
pi = {}
for s in mdp.S:
pi[s] = V[s][0]
V[s] = V[s][1]
return pi, V