def example_6_6():
fig, ax = plt.subplots()
fig.suptitle(f'Example 6.6 (Averaged over {EX_6_6_N_SEEDS} seeds)')
ax.set_xlabel('Episodes')
ax.set_ylabel(
f'(Average of last {EX_6_6_N_AVG}) sum of rewards during episodes')
ax.set_yticks(EX_6_6_YTICKS)
ax.set_ylim(bottom=min(EX_6_6_YTICKS))
n_ep = EX_6_6_N_EPS
env = TheCliff()
qlearning_alg = QLearning(env,
step_size=EX_6_5_STEP_SIZE,
gamma=UNDISCOUNTED,
eps=EX_6_5_EPS)
sarsa_alg = Sarsa(env,
step_size=EX_6_5_STEP_SIZE,
gamma=UNDISCOUNTED,
eps=EX_6_5_EPS)
qlearning_rew = np.zeros(n_ep)
sarsa_rew = np.zeros(n_ep)
for seed in range(EX_6_6_N_SEEDS):
print(f"seed={seed}")
qlearning_alg.seed(seed)
qlearning_rew += qlearning_alg.q_learning(n_ep)
sarsa_alg.seed(seed)
sarsa_rew += sarsa_alg.on_policy_td_control(n_ep, rews=True)
plt.plot(smooth_rewards(qlearning_rew / EX_6_6_N_SEEDS, EX_6_6_N_AVG),
color='r',
label='Q learning')
plt.plot(smooth_rewards(sarsa_rew / EX_6_6_N_SEEDS, EX_6_6_N_AVG),
color='b',
label='Sarsa')
plt.legend()
plt.savefig('example6.6.png')
plt.show()
def plot_sarsa(ax,
n_ep,
label=None,
diags=False,
stay=False,
stoch=False,
seed=0):
env = WindyGridworld(diags, stay, stoch)
alg = Sarsa(env,
step_size=EX_6_5_STEP_SIZE,
gamma=UNDISCOUNTED,
eps=EX_6_5_EPS)
alg.seed(seed)
kwargs = {"label": label} if label else {}
plt.plot(alg.on_policy_td_control(n_ep), **kwargs)
# Randomly locate the food on the barn.
amount_food = randint(max_size / 2, 2 * max_size)
food = []
while len(food) < amount_food:
# Add a new piece of food.
food.append((randint(0, max_size-1), randint(0, max_size-1)))
# Ensure uniqueness.
food = list(set(food))
# Start the algorithm.
sarsa = Sarsa(BarnState((0,0), food, max_size), epsilon=epsilon, alpha=alpha, gamma=gamma)
sarsa.seed(int(100 * time.time()))
# keep track of how much do we move the q.
track = []
for it in range(1, max_iters + 1):
if it % 10 == 0:
print "Scenario %d: %d/%d\r" % (n, it, max_iters) ,
sys.stdout.flush()
history, corrections = sarsa.iterate()
track.append(numpy.sqrt(sum(map(lambda x: x*x, corrections))))
# We're just selecting nice places to evaluate the current policy and create a picture.
if (it % 10 ** int(log10(it)) == 0) and (it / 10 ** int(log10(it)) in [1, 2, 4, 8]):