5 | 0.25355 | 1 | 0.254
6 | 0.10478 | 0 | 0.345
7 | 0.09657 | 0 | 0.442
8 | 0.03656 | 0 | 0.478
9 | 0.02772 | 0 | 0.506
10 | 0.01111 | 0 | 0.517
11 | 0.00735 | 0 | 0.524
12 | 0.00310 | 0 | 0.527
13 | 0.00190 | 0 | 0.529
14 | 0.00083 | 0 | 0.530
15 | 0.00049 | 0 | 0.531
16 | 0.00022 | 0 | 0.531
17 | 0.00013 | 0 | 0.531
18 | 0.00006 | 0 | 0.531
19 | 0.00003 | 0 | 0.531"""
Vs_VI, pis_VI = value_iteration(mdp, gamma=GAMMA, nIt=20, grade_print=make_grader(expected_output))
for (V, pi) in zip(Vs_VI[:10], pis_VI[:10]):
plt.figure(figsize=(3,3))
plt.imshow(V.reshape(4,4), cmap='gray', interpolation='none', clim=(0,1))
ax = plt.gca()
ax.set_xticks(np.arange(4)-.5)
ax.set_yticks(np.arange(4)-.5)
ax.set_xticklabels([])
ax.set_yticklabels([])
Y, X = np.mgrid[0:4, 0:4]
a2uv = {0: (-1, 0), 1:(0, -1), 2:(1,0), 3:(-1, 0)}
Pi = pi.reshape(4,4)
for y in range(4):
for x in range(4):
a = Pi[y, x]
expected_output = """Iteration | # chg actions | V[0]
----------+---------------+---------
0 | 1 | -0.00000
1 | 9 | 0.00000
2 | 2 | 0.39785
3 | 1 | 0.45546
4 | 0 | 0.53118
5 | 0 | 0.53118
6 | 0 | 0.53118
7 | 0 | 0.53118
8 | 0 | 0.53118
9 | 0 | 0.53118
10 | 0 | 0.53118
11 | 0 | 0.53118
12 | 0 | 0.53118
13 | 0 | 0.53118
14 | 0 | 0.53118
15 | 0 | 0.53118
16 | 0 | 0.53118
17 | 0 | 0.53118
18 | 0 | 0.53118
19 | 0 | 0.53118"""
Vs_PI, pis_PI = policy_iteration(mdp,
gamma=0.95,
nIt=20,
grade_print=make_grader(expected_output))
plt.plot(Vs_PI)
def compute_vpi(pi, mdp, gamma):
# use pi[state] to access the action that's prescribed by this policy
V = np.ones(mdp.nS) # REPLACE THIS LINE WITH YOUR CODE
A = np.zeros((mdp.nS,mdp.nS))
B = np.zeros(mdp.nS)
for s in range(mdp.nS):
a = pi[s]
# p = mdp.P[s][a][0]
# ns = mdp.P[s][a][1]
# r = mdp.P[s][a][2]
for p, ns, r in mdp.P[s][pi[s]]:
A[s][ns] += GAMMA*p
B[s] += p*r
A = A - np.eye(mdp.nS)
V = -np.linalg.solve(A,B)
#V[s] = mdp.P[s][pi[s]][0]*(mdp.P[s][pi[s]][2] + GAMMA*V[mdp.P[s][a][1]])
return V
def compute_qpi(vpi, mdp, gamma):
Qpi = np.zeros([mdp.nS, mdp.nA]) # REPLACE THIS LINE WITH YOUR CODE
for s in range(mdp.nS):
for a in range(mdp.nA):
for p,ns,r in mdp.P[s][a]:
Qpi[s,a] += p*r + gamma*p*vpi[ns]
return Qpi
expected_Qpi = np.array([[ 0.38 , 3.135, 1.14 , 0.095],
[ 0.57 , 3.99 , 2.09 , 0.95 ],
[ 1.52 , 4.94 , 3.04 , 1.9 ],
[ 2.47 , 5.795, 3.23 , 2.755],
[ 3.8 , 6.935, 4.56 , 0.855],
[ 4.75 , 4.75 , 4.75 , 4.75 ],
[ 4.94 , 8.74 , 6.46 , 2.66 ],
[ 6.65 , 6.65 , 6.65 , 6.65 ],
[ 7.6 , 10.735, 8.36 , 4.655],
[ 7.79 , 11.59 , 9.31 , 5.51 ],
[ 8.74 , 12.54 , 10.26 , 6.46 ],
[ 10.45 , 10.45 , 10.45 , 10.45 ],
[ 11.4 , 11.4 , 11.4 , 11.4 ],
[ 11.21 , 12.35 , 12.73 , 9.31 ],
[ 12.16 , 13.4 , 14.48 , 10.36 ],
[ 14.25 , 14.25 , 14.25 , 14.25 ]])
Qpi = compute_qpi(np.arange(mdp.nS), mdp, gamma=0.95)
if np.all(np.isclose(expected_Qpi, Qpi, atol=1e-4)):
print("Test passed")
else:
print("Expected: ", expected_Qpi)
print("Actual: ", Qpi)
def policy_iteration(mdp, gamma, nIt, grade_print=print):
Vs = []
pis = []
pi_prev = np.zeros(mdp.nS,dtype='int')
pis.append(pi_prev)
grade_print("Iteration | # chg actions | V[0]")
grade_print("----------+---------------+---------")
for it in range(nIt):
# YOUR CODE HERE
# you need to compute qpi which is the state-action values for current pi
vpi = compute_vpi(pi_prev, mdp, gamma)
qpi = compute_qpi(vpi, mdp, gamma)
pi = qpi.argmax(axis=1)
grade_print("%4i | %6i | %6.5f"%(it, (pi != pi_prev).sum(), vpi[0]))
Vs.append(vpi)
pis.append(pi)
pi_prev = pi
return Vs, pis
expected_output = """Iteration | # chg actions | V[0]
----------+---------------+---------
0 | 1 | -0.00000
1 | 9 | 0.00000
2 | 2 | 0.39785
3 | 1 | 0.45546
4 | 0 | 0.53118
5 | 0 | 0.53118
6 | 0 | 0.53118
7 | 0 | 0.53118
8 | 0 | 0.53118
9 | 0 | 0.53118
10 | 0 | 0.53118
11 | 0 | 0.53118
12 | 0 | 0.53118
13 | 0 | 0.53118
14 | 0 | 0.53118
15 | 0 | 0.53118
16 | 0 | 0.53118
17 | 0 | 0.53118
18 | 0 | 0.53118
19 | 0 | 0.53118"""
Vs_PI, pis_PI = policy_iteration(mdp, gamma=0.95, nIt=20, grade_print=make_grader(expected_output))
plt.plot(Vs_PI);