def __init__(self,
env,
t,
alphas=None,
betas=None,
thres=0.5,
iters_per_arm=100):
self.env = snapshot(env, t)
self.t = t
self.arms = two_cycles(self.env, t)
self.n_arms = len(self.arms)
self.iters_per_arm = iters_per_arm
self.thres = thres
# Prior successes
if alphas is None:
self.alphas = np.ones(self.n_arms)
else:
self.alphas = alphas
# Prior failures
if betas is None:
self.betas = np.ones(self.n_arms)
else:
self.betas = alphas
self.s = np.zeros(self.n_arms) # Successes
self.f = np.zeros(self.n_arms) # Failures
self.r = np.zeros(self.n_arms) # Rewards
self.n = np.zeros(self.n_arms) # Visits
def __init__(self,
env,
t,
n_iters=10,
method="cycles",
mix=0.5,
none_prob=None):
self.env = snapshot(env, t)
self.t = t
self.n_iters = n_iters
self.method = method
if method == "cycles":
self.arms = list(
map(lambda x: tuple(sorted(x)), two_cycles(self.env,
t))) + [None]
elif method == "pairs":
self.arms = self.env.get_living(self.t) + [None]
else:
raise ValueError("Unknown optimal simulation method.")
self.n_arms = len(self.arms)
self.rnn = torch.load("results/policy_function_lstm")
self.mix = mix
self.none_prob = none_prob
def test_two_cycles_all(env):
cycles = two_cycles(env, 4)
for i, j in cycles:
assert env.has_edge(i, j)
assert env.has_edge(j, i)
assert env.node[i]["entry"] <= env.node[j]["death"]
assert env.node[j]["entry"] <= env.node[i]["death"]
assert env.node[i]["d_blood"] == 0 or \
env.node[j]["p_blood"] == 3 or \
env.node[i]["d_blood"] == env.node[j]["p_blood"]
def __init__(self, env, t, priors, prior_counts, algo="opt"):
self.env = snapshot(env, t)
self.t = t
self.arms = two_cycles(self.env, t)
self.n_arms = len(self.arms)
self.successes = np.zeros(self.n_arms)
self.solver = optimal
self.horizons = [1, 5, 10, 20]
self.successes = prior_counts * get_cycle_probabilities(
self.env.get_living(t), self.arms, priors) + 1e-8
def __init__(self, env, t, gamma=.1, iters_per_arm=100, thres=0.5):
self.env = env
self.t = t
self.arms = two_cycles(self.env, t)
self.n_arms = len(self.arms)
self.w = np.ones(self.n_arms)
self.p = np.full_like(self.w, fill_value=1/self.n_arms)
self.gamma = gamma
self.r = np.zeros(self.n_arms)
self.n = np.zeros(self.n_arms)
self.iters_per_arm = iters_per_arm
self.thres = thres
def __init__(self, env, t, c=2, iters_per_arm=100, thres=0.5):
self.env = snapshot(env, t)
self.t = t
self.arms = two_cycles(self.env, t)
self.n_arms = len(self.arms)
self.c = c
self.r = np.zeros(self.n_arms) # Rewards
self.n = np.zeros(self.n_arms) # Visits
self.iters_per_arm = iters_per_arm
self.thres = thres
def __init__(self, env, t, method="cycles"):
self.env = snapshot(env, t)
self.t = t
self.method = method
if method == "cycles":
self.arms = list(
map(lambda x: tuple(sorted(x)), two_cycles(self.env,
t))) + [None]
elif method == "pairs":
self.arms = self.env.get_living(self.t) + [None]
else:
raise ValueError("Unknown optimal simulation method.")
self.n_arms = len(self.arms)
def __init__(self,
env,
t: int,
gamma: float = .1,
iters_per_arm: int = 100,
max_match: int = 5):
self.env = env
self.t = t
self.max_match = max_match
self.cycles = two_cycles(env, t) + [()]
self.arms = get_arm_matrix(self.cycles, max_match)
self.n_arms = len(self.arms)
self.h = max(2, int(geom(env.death_rate).ppf(.9)))
self.w = np.ones(self.n_arms)
self.p = np.full_like(self.w, fill_value=1 / self.n_arms)
self.q = np.full_like(self.w, fill_value=1 / self.n_arms)
self.mu = np.full_like(self.w, fill_value=1 / self.n_arms)
self.gamma = gamma
self.iters_per_arm = iters_per_arm
opt = optimal(env)
gre = greedy(env)
o = get_n_matched(opt["matched"], 0, env.time_length)
g = get_n_matched(gre["matched"], 0, env.time_length)
rewards = np.zeros(env.time_length)
#%%
np.random.seed(clock_seed())
for t in range(env.time_length):
probs, count = evaluate_policy(net, env, t, dtype="numpy")
for i in range(count):
probs, _ = evaluate_policy(net, env, t, dtype="numpy")
cycles = two_cycles(env, t)
if len(cycles) == 0:
break
elif len(cycles) == 1:
res = cycles.pop()
else:
sim = MonteCarlo(env, t, probs, n_prior)
res = sim.simulate(n_sims)
env.removed_container[t].update(res)
probs, count = evaluate_policy(net, env, t, dtype="numpy")
rewards[t] = len(env.removed_container[t])
print(t, np.mean(rewards[:t + 1]), np.mean(g[:t + 1]),
np.mean(o[:t + 1]))
