def descendants(self):
"""
All child nodes and all their child nodes.
>>> from anytree import Node
>>> udo = Node("Udo")
>>> marc = Node("Marc", parent=udo)
>>> lian = Node("Lian", parent=marc)
>>> loui = Node("Loui", parent=marc)
>>> soe = Node("Soe", parent=lian)
>>> udo.descendants
(Node('/Udo/Marc'), Node('/Udo/Marc/Lian'), Node('/Udo/Marc/Lian/Soe'), Node('/Udo/Marc/Loui'))
>>> marc.descendants
(Node('/Udo/Marc/Lian'), Node('/Udo/Marc/Lian/Soe'), Node('/Udo/Marc/Loui'))
>>> lian.descendants
(Node('/Udo/Marc/Lian/Soe'),)
"""
return tuple(PreOrderIter(self))[1:]
def prepare_archetype(self, a, archetypes):
a.current = a.id == getattr(self, 'archetype', {}).get('id', None)
if a.get('all_num_decks') is not None and a.get(
'season_num_decks') is not None:
a.all_show_record = a.get('all_wins') or a.get(
'all_draws') or a.get('all_losses')
a.season_show_record = a.get('season_wins') or a.get(
'season_draws') or a.get('season_losses')
a.show_matchups = a.all_show_record
a.url = '/archetypes/{id}/'.format(id=a.id)
a.best_decks = Container({'decks': []})
n = 3
while len(a.best_decks.decks) == 0 and n >= 0:
for d in a.get('decks', []):
if d.get('stars_safe', '').count('★') >= n:
a.best_decks.decks.append(d)
n -= 1
a.show_best_decks = len(a.decks) != len(a.best_decks.decks)
counter = Counter()
a.cards = []
a.most_common_cards = []
for d in a.get('decks', []):
a.cards += d.maindeck + d.sideboard
for c in d.maindeck:
if not c['card'].type.startswith('Basic Land'):
counter[c['name']] += c['n']
most_common_cards = counter.most_common(NUM_MOST_COMMON_CARDS_TO_LIST)
cs = oracle.cards_by_name()
for v in most_common_cards:
self.prepare_card(cs[v[0]])
a.most_common_cards.append(cs[v[0]])
a.has_most_common_cards = len(a.most_common_cards) > 0
a.archetype_tree = PreOrderIter(a)
for r in a.archetype_tree:
# Prune branches we don't want to show
if r.id not in [a.id for a in archetypes]:
r.parent = None
r['url'] = '/archetypes/{id}/'.format(id=r['id'])
# It perplexes me that this is necessary. It's something to do with the way NodeMixin magic works. Mustache doesn't like it.
r['depth'] = r.depth
def prepare_archetype(self, a: archetype.Archetype, archetypes: List[archetype.Archetype]) -> None:
a.current = a.id == getattr(self, 'archetype', {}).get('id', None)
a.show_record = a.get('num_decks') is not None and (a.get('wins') or a.get('draws') or a.get('losses'))
counter = Counter() # type: ignore
a.cards = []
a.most_common_cards = []
# Make a pass, collecting card counts for all decks and for tournament decks
for d in a.get('decks', []):
a.cards += d.maindeck + d.sideboard
for c in d.maindeck:
if not c.card.type_line.startswith('Basic Land'):
counter[c['name']] += c['n']
most_common_cards = counter.most_common(prepare.NUM_MOST_COMMON_CARDS_TO_LIST)
cs = oracle.cards_by_name()
for v in most_common_cards:
prepare.prepare_card(cs[v[0]], getattr(self, 'tournament_only', False))
a.most_common_cards.append(cs[v[0]])
a.has_most_common_cards = len(a.most_common_cards) > 0
for b in [b for b in PreOrderIter(a) if b.id in [a.id for a in archetypes]]:
b['url'] = url_for('.archetype', archetype_id=b['id'])
# It perplexes me that this is necessary. It's something to do with the way NodeMixin magic works. Mustache doesn't like it.
b['depth'] = b.depth
def MCTS(self, bel, d):
h = self.T.name
for a in range(0, self.model.acts):
tmp = Node(self.T.name + str(a),
parent=self.T,
value=self.Q0,
count=self.N0)
numLoops = 100
for i in range(0, numLoops):
s = np.random.choice([i for i in range(0, self.model.N)], p=bel)
self.simulate(s, h, d)
#RenderTreeGraph(self.T).to_picture('tree1.png');
#print(RenderTree(self.T));
QH = [0] * self.model.acts
for a in range(0, self.model.acts):
QH[a] = [
node.value for node in PreOrderIter(
self.T, filter_=lambda n: n.name == h + str(a))
][0]
act = np.argmax([QH[a] for a in range(0, self.model.acts)])
return [act, QH[act]]
def prepare_archetype(self, a, archetypes):
num_most_common_cards_to_list = 10
a.current = a.id == getattr(self, 'archetype', {}).get('id', None)
if a.get('all') and a.get('season'):
a.all.show_record = a.all.get('wins') or a.all.get(
'draws') or a.all.get('losses')
a.season.show_record = a.season.get('wins') or a.season.get(
'draws') or a.season.get('losses')
a.url = url_for('archetype', archetype_id=a.id)
a.best_decks = Container({'decks': []})
n = 3
while len(a.best_decks.decks) == 0 and n >= 0:
for d in a.get('decks', []):
if d.get('stars_safe', '').count('★') >= n:
a.best_decks.decks.append(d)
n -= 1
counter = Counter()
a.cards = []
a.most_common_cards = []
for d in a.get('decks', []):
a.cards += d.maindeck + d.sideboard
for c in d.maindeck:
if not c['card'].type.startswith('Basic Land'):
counter[c['name']] += c['n']
most_common_cards = counter.most_common(num_most_common_cards_to_list)
cs = oracle.cards_by_name()
for v in most_common_cards:
self.prepare_card(cs[v[0]])
a.most_common_cards.append(cs[v[0]])
a.archetype_tree = PreOrderIter(a)
for r in a.archetype_tree:
# Prune branches we don't want to show
if r.id not in [a.id for a in archetypes]:
r.parent = None
r['url'] = url_for('archetype', archetype_id=r['id'])
# It perplexes me that this is necessary. It's something to do with the way NodeMixin magic works. Mustache doesn't like it.
r['depth'] = r.depth
def get_node(self, node_name):
for node in PreOrderIter(self.get_root()):
if node.name == node_name:
return node
raise ValueError("Node {} not found".format(node_name))
def testMCTSSim():
trails = 10
trailLength = 100
allReward = np.zeros(shape=(trails, trailLength)).tolist()
for count in range(0, trails):
if (trails == 1):
fig, ax = plt.subplots()
totalReward = 0
a = OnlineSolver()
b = np.random.rand(a.model.N).tolist()
x = np.random.randint(0, a.model.N)
b[x] = 3
b = a.normalize(b)
for step in range(0, trailLength):
if (trails == 1):
ax.cla()
ax.plot(b, linewidth=4)
ax.scatter(x, .4, s=150, c='r')
ax.set_ylim([0, .5])
ax.set_title('POMCP Belief')
plt.pause(0.1)
[act, u] = a.MCTS(b, 3)
totalReward += a.model.R[act][x]
x = np.random.choice([i for i in range(0, a.model.N)],
p=a.model.px[act][x])
ztrial = [0] * len(a.model.pz)
for i in range(0, len(a.model.pz)):
ztrial[i] = a.model.pz[i][x]
z = ztrial.index(max(ztrial))
b = a.beliefUpdate(b, act, z)
a.T = [
node for node in PreOrderIter(
a.T,
filter_=lambda n: n.name == a.T.name + str(act) + str(z))
][0]
#RenderTreeGraph(a.T).to_picture('tree2.png');
a.T.parent = None
#print(a.T);
#RenderTreeGraph(a.T).to_picture('tree1.png');
allReward[count][step] = totalReward
print(allReward[count][-1])
averageAllReward = [0] * trailLength
for i in range(0, trails):
for j in range(0, trailLength):
averageAllReward[j] += allReward[i][j] / trails
allSigma = [0] * trailLength
for i in range(0, trailLength):
suma = 0
for j in range(0, trails):
suma += (allReward[j][i] - averageAllReward[i])**2
allSigma[i] = np.sqrt(suma / trails)
UpperBound = [0] * trailLength
LowerBound = [0] * trailLength
for i in range(0, trailLength):
UpperBound[i] = averageAllReward[i] + allSigma[i]
LowerBound[i] = averageAllReward[i] - allSigma[i]
x = [i for i in range(0, trailLength)]
plt.figure()
plt.plot(x, averageAllReward, 'g')
plt.plot(x, UpperBound, 'g--')
plt.plot(x, LowerBound, 'g--')
plt.fill_between(x, LowerBound, UpperBound, color='g', alpha=0.25)
plt.xlabel('Time Step')
plt.ylabel('Accumlated Reward')
plt.title('Average Accumulated Rewards over Time for: ' + str(trails) +
' simulations')
plt.show()
def testMCTSSim2D():
trails = 10
trailLength = 100
allReward = np.zeros(shape=(trails, trailLength)).tolist()
random = False
for count in range(0, trails):
'''
if(trails == 1):
fig,ax = plt.subplots();
'''
totalReward = 0
a = OnlineSolver()
x1 = np.random.randint(-5, 5)
x2 = np.random.randint(-5, 5)
x = [x1, x2]
b = GM()
b.addG(Gaussian(x, [[1, 0], [0, 1]], 1))
for step in range(0, trailLength):
'''
if(trails == 1):
ax.cla();
ax.plot(b,linewidth=4);
ax.scatter(x,.4,s=150,c='r');
ax.set_ylim([0,.5]);
ax.set_title('POMCP Belief');
plt.pause(0.1);
'''
if (random):
act = np.random.randint(0, 5)
else:
[act, u] = a.MCTS(b, 2)
totalReward += a.model.r[act].pointEval(x)
#x = np.random.choice([i for i in range(0,a.model.N)],p=a.model.px[act][x]);
tmpGM = GM()
tmpGM.addG(
Gaussian((np.array(x) + np.array(a.model.delA[act])).tolist(),
a.model.delAVar, 1))
x = tmpGM.sample(1)[0]
ztrial = [0] * len(a.model.pz)
for i in range(0, len(a.model.pz)):
ztrial[i] = a.model.pz[i].pointEval(x)
z = ztrial.index(max(ztrial))
b = a.beliefUpdate(b, act, z, a.model)
if (not random):
#RenderTreeGraph(a.T).to_picture('tree2.png');
a.T = [
node for node in PreOrderIter(a.T,
filter_=lambda n: n.name == a
.T.name + str(act) + str(z))
][0]
a.T.parent = None
#print(a.T);
#RenderTreeGraph(a.T).to_picture('tree1.png');
allReward[count][step] = totalReward
print(allReward[count][-1])
averageAllReward = [0] * trailLength
for i in range(0, trails):
for j in range(0, trailLength):
averageAllReward[j] += allReward[i][j] / trails
allSigma = [0] * trailLength
for i in range(0, trailLength):
suma = 0
for j in range(0, trails):
suma += (allReward[j][i] - averageAllReward[i])**2
allSigma[i] = np.sqrt(suma / trails)
UpperBound = [0] * trailLength
LowerBound = [0] * trailLength
for i in range(0, trailLength):
UpperBound[i] = averageAllReward[i] + allSigma[i]
LowerBound[i] = averageAllReward[i] - allSigma[i]
x = [i for i in range(0, trailLength)]
plt.figure()
plt.plot(x, averageAllReward, 'g')
plt.plot(x, UpperBound, 'g--')
plt.plot(x, LowerBound, 'g--')
plt.fill_between(x, LowerBound, UpperBound, color='g', alpha=0.25)
plt.xlabel('Time Step')
plt.ylabel('Accumlated Reward')
plt.title('Average Accumulated Rewards over Time for: ' + str(trails) +
' simulations')
plt.show()
def simulate(self, s, h, d):
if (d == 0):
return 0
if (len([
node
for node in PreOrderIter(self.T, filter_=lambda n: n.name == h)
]) == 0):
newRoot = [
node for node in PreOrderIter(
self.T, filter_=lambda n: n.name == h[0:len(h) - 2])
][0]
newName = h[0:len(h) - 2]
for a in range(0, self.model.acts):
if (len([
node for node in PreOrderIter(
self.T,
filter_=lambda n: n.name == newName + str(a))
]) == 0):
tmp = Node(h + str(a),
parent=newRoot,
value=self.Q0,
count=self.N0)
for o in range(0, self.model.acts):
if (len([
node for node in PreOrderIter(
self.T,
filter_=lambda n: n.name == newName + str(
a) + str(o))
]) == 0):
tmp2 = Node(h + str(a) + str(o),
parent=tmp,
value=self.Q0,
count=self.N0)
#tmp = Node(h,parent=newRoot,value = self.Q0,count=self.N0);
return self.getRolloutReward(s, d)
else:
newRoot = [
node
for node in PreOrderIter(self.T, filter_=lambda n: n.name == h)
][0]
newName = h
for a in range(0, self.model.acts):
if (len([
node for node in PreOrderIter(
self.T,
filter_=lambda n: n.name == newName + str(a))
]) == 0):
tmp = Node(h + str(a),
parent=newRoot,
value=self.Q0,
count=self.N0)
for o in range(0, self.model.acts):
if (len([
node for node in PreOrderIter(
self.T,
filter_=lambda n: n.name == newName + str(
a) + str(o))
]) == 0):
tmp2 = Node(h + str(a) + str(o),
parent=tmp,
value=self.Q0,
count=self.N0)
QH = [0] * self.model.acts
NH = [0] * self.model.acts
NodeH = [0] * self.model.acts
#print([node.name for node in PreOrderIter(self.T)])
#RenderTreeGraph(self.T).to_picture('tree1.png');
for a in range(0, self.model.acts):
#print(h,a);
QH[a] = [
node.value for node in PreOrderIter(
self.T, filter_=lambda n: n.name == h + str(a))
][0]
NH[a] = [
node.count for node in PreOrderIter(
self.T, filter_=lambda n: n.name == h + str(a))
][0]
NodeH[a] = [
node for node in PreOrderIter(
self.T, filter_=lambda n: n.name == h + str(a))
][0]
aprime = np.argmax([
QH[a] + self.exploreParam * np.sqrt(np.log(sum(NH) / NH[a]))
for a in range(0, self.model.acts)
])
[sprime, o, r] = self.generate(s, aprime)
q = r + self.model.discount * self.simulate(
sprime, h + str(aprime) + str(o), d - 1)
NodeH[aprime].count += 1
NodeH[aprime].value += (q - QH[a]) / NH[a]
return q
