def _add_to_cluster(self, cluster, _doc): super(WavgNetCONFIRM, self)._add_to_cluster(cluster, _doc) # competitive stage similarities = self._cluster_sim_scores(_doc) idx = utils.argmax(similarities) del similarities[idx] if similarities: idx2 = utils.argmax(similarities) if idx2 <= idx: idx2 += 1 sim_vec2 = self.clusters[idx2].center.similarity_vector(_doc) self.clusters[idx2].network.learn(sim_vec2, 0.2)
def best_policy(mdp, U):
"""Given an MDP and a utility function U, determine the best policy,
as a mapping from state to action. (Equation 17.4)"""
pi = {}
for s in mdp.states:
pi[s] = argmax(mdp.actions(s), key=lambda a: expected_utility(a, s, U, mdp))
return pi
def genetic_algorithm(problem, population, fitness_fn, ngen=1000, pmut=0.1):
"[Fig. 4.8]"
#MAX = 0
for i in range(ngen):
new_population = []
'''
print i, '------------'
print ' ', MAX
for p in population:
print problem.value(p)
if problem.value(p) > MAX:
MAX = problem.value(p)
'''
for p in population:
fitnesses = map(fitness_fn, population)
s1, s2 = weighted_sample_with_replacement(population, fitnesses, 2)
p1 = copy.copy(problem)
p1.set_state(s1)
p2 = copy.copy(problem)
p2.set_state(s2)
child = p1.mate(p2)
child.mutate(pmut)
new_population.append(child.initial)
population = new_population
return utils.argmax(population, fitness_fn)
def update(self,x,y):
"""
updates the ORT
- x : list of k covariates (k x 1)
- y : response (scalar)
usage:
ort.update(x,y)
"""
k = self.__poisson(1)
if k == 0:
self.__updateOOBE(x,y)
else:
for u in xrange(k):
self.__age += 1
(j,depth) = self.__findLeaf(x,self.tree)
j.elem.update(x,y)
#if j.elem.numSamplesSeen > self.minSamples and depth < self.maxDepth: # FIXME: which is the correct approach?
if j.elem.stats.n > self.minSamples and depth < self.maxDepth:
g = self.__gains(j.elem)
if any([ gg >= self.minGain for gg in g ]):
bestTest = j.elem.tests[argmax(g)]
j.elem.updateSplit(bestTest.dim,bestTest.loc)
j.updateChildren( Tree(Elem(self.param)), Tree(Elem(self.param)) )
j.left.elem.stats = bestTest.statsL
j.right.elem.stats = bestTest.statsR
j.elem.reset()
def genetic_algorithm_stepwise(population):
root.title('Genetic Algorithm')
for generation in range(ngen):
# generating new population after selecting, recombining and mutating the existing population
population = [search.mutate(search.recombine(*search.select(2, population, fitness_fn)), gene_pool, mutation_rate) for i in range(len(population))]
# genome with the highest fitness in the current generation
current_best = ''.join(argmax(population, key=fitness_fn))
# collecting first few examples from the current population
members = [''.join(x) for x in population][:48]
# clear the canvas
canvas.delete('all')
# displays current best on top of the screen
canvas.create_text(canvas_width / 2, 40, fill=p_blue, font='Consolas 46 bold', text=current_best)
# displaying a part of the population on the screen
for i in range(len(members) // 3):
canvas.create_text((canvas_width * .175), (canvas_height * .25 + (25 * i)), fill=lp_blue, font='Consolas 16', text=members[3 * i])
canvas.create_text((canvas_width * .500), (canvas_height * .25 + (25 * i)), fill=lp_blue, font='Consolas 16', text=members[3 * i + 1])
canvas.create_text((canvas_width * .825), (canvas_height * .25 + (25 * i)), fill=lp_blue, font='Consolas 16', text=members[3 * i + 2])
# displays current generation number
canvas.create_text((canvas_width * .5), (canvas_height * 0.95), fill=p_blue, font='Consolas 18 bold', text=f'Generation {generation}')
# displays blue bar that indicates current maximum fitness compared to maximum possible fitness
scaling_factor = fitness_fn(current_best) / len(target)
canvas.create_rectangle(canvas_width * 0.1, 90, canvas_width * 0.9, 100, outline=p_blue)
canvas.create_rectangle(canvas_width * 0.1, 90, canvas_width * 0.1 + scaling_factor * canvas_width * 0.8, 100, fill=lp_blue)
canvas.update()
# checks for completion
fittest_individual = search.fitness_threshold(fitness_fn, f_thres, population)
if fittest_individual:
break
def WalkSAT(clauses, p=0.5, max_flips=10000):
"""Checks for satisfiability of all clauses by randomly flipping values of variables
"""
# Set of all symbols in all clauses
symbols = set(sym for clause in clauses for sym in prop_symbols(clause))
# model is a random assignment of true/false to the symbols in clauses
model = {s: random.choice([True, False]) for s in symbols}
for i in range(max_flips):
satisfied, unsatisfied = [], []
for clause in clauses:
(satisfied if pl_true(clause, model) else unsatisfied).append(clause)
if not unsatisfied: # if model satisfies all the clauses
return model
clause = random.choice(unsatisfied)
if probability(p):
sym = random.choice(prop_symbols(clause))
else:
# Flip the symbol in clause that maximizes number of sat. clauses
def sat_count(sym):
# Return the the number of clauses satisfied after flipping the symbol.
model[sym] = not model[sym]
count = len([clause for clause in clauses if pl_true(clause, model)])
model[sym] = not model[sym]
return count
sym = argmax(prop_symbols(clause), key=sat_count)
model[sym] = not model[sym]
# If no solution is found within the flip limit, we return failure
return None
def determine_optimum_variants(unit1, unit2):
"""Determines the optimum variants between two units."""
# TODO - improve performance by considering variants (1,1) (1, 2) and (2,1)
# as equivalent.
outcomes = defaultdict(dict)
for v1 in MeleeRangedStrategy.VARIANTS:
if not MeleeRangedStrategy.is_compatible(unit1, v1):
continue
unit1.strategy = MeleeRangedStrategy(unit1, v1)
for v2 in MeleeRangedStrategy.VARIANTS:
if not MeleeRangedStrategy.is_compatible(unit2, v2):
continue
unit2.strategy = MeleeRangedStrategy(unit2, v2)
turn_order = (unit1, unit2)
game_state = AveragingVersusGameState(turn_order, verbosity=0)
game_state.run_combat()
outcomes[v1][v2] = game_state.hp_delta
# What's your best strategy?
unit_1_strategies = { v1: min(outcomes[v1].values()) for v1 in outcomes }
unit1_strategy = utils.argmax(unit_1_strategies)
unit2_strategy = utils.argmin(outcomes[unit1_strategy])
# for v1 in outcomes:
# for v2, hp_delta in sorted(outcomes[v1].items()):
# print '(%d, %d) => %+.2f' % (v1, v2, hp_delta)
# print '%s\'s strategy: %s' % (unit1, unit1_strategy)
# print '%s\'s strategy: %s' % (unit2, unit2_strategy)
return (unit1_strategy, unit2_strategy)
def alphabeta_full_search(state, game):
"""Search game to determine best action; use alpha-beta pruning.
As in [Fig. 6.7], this version searches all the way to the leaves."""
player = game.to_move(state)
def max_value(state, alpha, beta):
if game.terminal_test(state):
return game.utility(state, player)
v = -infinity
for (a, s) in game.successors(state):
v = max(v, min_value(s, alpha, beta))
if v >= beta:
return v
alpha = max(alpha, v)
return v
def min_value(state, alpha, beta):
if game.terminal_test(state):
return game.utility(state, player)
v = infinity
for (a, s) in game.successors(state):
v = min(v, max_value(s, alpha, beta))
if v <= alpha:
return v
beta = min(beta, v)
return v
# Body of alphabeta_search starts here:
action, state = argmax(game.successors(state),
lambda ((a, s)): min_value(s, -infinity, infinity))
return action
def minimax_decision(state, game):
"""Given a state in a game, calculate the best move by searching
forward all the way to the terminal states. [Fig. 6.4]"""
player = game.to_move(state)
def max_value(state):
if game.terminal_test(state):
return game.utility(state, player)
v = -infinity
for (_, s) in game.successors(state):
v = max(v, min_value(s))
return v
def min_value(state):
if game.terminal_test(state):
return game.utility(state, player)
v = infinity
for (_, s) in game.successors(state):
v = min(v, max_value(s))
return v
# Body of minimax_decision starts here:
action, state = argmax(game.successors(state), lambda ((a, s)): min_value(s))
return action
def minimax_decision(state, game):
"""Given a state in a game, calculate the best move by searching
forward all the way to the terminal states. [Figure 5.3]"""
player = game.to_move(state)
def max_value(state):
if game.terminal_test(state):
return game.utility(state, player)
v = -infinity
for a in game.actions(state):
v = max(v, min_value(game.result(state, a)))
return v
def min_value(state):
if game.terminal_test(state):
return game.utility(state, player)
v = infinity
for a in game.actions(state):
v = min(v, max_value(game.result(state, a)))
return v
# Body of minimax_decision:
return argmax(game.actions(state),
key=lambda a: min_value(game.result(state, a)))
def actions(self, state):
search_list = [c for c in self.decoder.chardomain if c not in state]
target_list = [c for c in alphabet if c not in state.values()]
# Find the best charater to replace
plainchar = argmax(search_list, key=lambda c: self.decoder.P1[c])
for cipherchar in target_list:
yield (plainchar, cipherchar)
def _add_to_cluster(self, cluster, _doc): super(PushAwayCONFIRM, self)._add_to_cluster(cluster, _doc) sim_score = self._cached_most_similar_val margins = map(lambda x: sim_score - x if sim_score != x else 0, self._get_cached_sim_scores(_doc)) most_similar_cluster = self.clusters[utils.argmax(margins)] cluster.center.push_away(most_similar_cluster.center)
def generate(self, in_str):
dy.renew_cg()
in_seq = u.preprocess_seq(in_str, self.char2int)
output_str = ""
for probs in self._probs(in_seq):
char = u.argmax(probs)
output_str += self.int2char[char]
return output_str
def predict(example):
"""Predict the target value for example. Consider each possible value,
and pick the most likely by looking at each attribute independently."""
def class_probability(targetval):
return (target_dist[targetval] *
product(attr_dists[targetval, attr][example[attr]]
for attr in dataset.inputs))
return argmax(target_vals, key=class_probability)
def best_policy(mdp, U):
"""Given an MDP and a utility function U, determine the best policy,
as a mapping from state to action. (Equation 17.4)"""
pi = {}
for s in mdp.states:
pi[s] = argmax(mdp.actions(s),
key=lambda a: expected_utility(a, s, U, mdp))
return pi
def renderResults(name, results, test_x, activations_0, activations_1):
image = Image.new('RGB', (28 * 3 + 280, len(results) * 280))
draw = ImageDraw.Draw(image)
for index, result in enumerate(results):
renderImage(image,
"MNIST_data/" + str(argmax(test_x[index][0])) + ".png",
0 * 28, index * 280 + 0 * 32)
renderImage(image, "MNIST_data/" + str(argmax(result[0])) + ".png",
2 * 28, index * 280 + 0 * 32)
renderActivations(draw, activations_1[index][0], 3 * 28,
index * 280 + 0 * 32)
renderActivations(draw, activations_0[index][0], 3 * 28,
index * 280 + 0 * 32 + 14)
renderImage(image,
"MNIST_data/" + str(argmax(test_x[index][1])) + ".png",
0 * 28, index * 280 + 1 * 32)
renderImage(image, "MNIST_data/" + str(argmax(result[1])) + ".png",
2 * 28, index * 280 + 1 * 32)
renderActivations(draw, activations_1[index][1], 3 * 28,
index * 280 + 1 * 32)
renderActivations(draw, activations_0[index][1], 3 * 28,
index * 280 + 1 * 32 + 14)
renderImage(image,
"MNIST_data/" + str(argmax(test_x[index][2])) + ".png",
0 * 28, index * 280 + 2 * 32)
renderImage(image, "MNIST_data/" + str(argmax(result[2])) + ".png",
2 * 28, index * 280 + 2 * 32)
renderActivations(draw, activations_1[index][2], 3 * 28,
index * 280 + 2 * 32)
renderActivations(draw, activations_0[index][2], 3 * 28,
index * 280 + 2 * 32 + 14)
renderImage(image, "MNIST_data/" + "plus.png", 0 * 28,
index * 280 + 3 * 32)
renderImage(image, "MNIST_data/" + str(argmax(result[3])) + ".png",
2 * 28, index * 280 + 3 * 32)
renderActivations(draw, activations_1[index][3], 3 * 28,
index * 280 + 3 * 32)
renderActivations(draw, activations_0[index][3], 3 * 28,
index * 280 + 3 * 32 + 14)
renderImage(image, "MNIST_data/" + "equals.png", 0 * 28,
index * 280 + 4 * 32)
renderImage(image, "MNIST_data/" + str(argmax(result[4])) + ".png",
2 * 28, index * 280 + 4 * 32)
renderActivations(draw, activations_1[index][4], 3 * 28,
index * 280 + 4 * 32)
renderActivations(draw, activations_0[index][4], 3 * 28,
index * 280 + 4 * 32 + 14)
image.save(name + ".png")
def renderResults200(name, results, test_x, activations_0, activations_1):
image = Image.new('RGB', (28 * 3 + 800, len(results) * 180))
draw = ImageDraw.Draw(image)
for index, result in enumerate(results):
renderDigit(draw, test_x[index][0], 0 * 28, index * 180 + 0 * 36)
renderImage(image,
"MNIST_data/" + str(argmax(result[0][10:])) + ".png",
2 * 28, index * 180 + 0 * 36)
renderActivations200(draw, activations_0[index][0], 3 * 28,
index * 180 + 0 * 36)
renderActivations200(draw, activations_1[index][0], 3 * 28,
index * 180 + 0 * 36 + 16)
renderDigit(draw, test_x[index][1], 0 * 28, index * 180 + 1 * 36)
renderImage(image,
"MNIST_data/" + str(argmax(result[1][10:])) + ".png",
2 * 28, index * 180 + 1 * 36)
renderActivations200(draw, activations_0[index][1], 3 * 28,
index * 180 + 1 * 36)
renderActivations200(draw, activations_1[index][1], 3 * 28,
index * 180 + 1 * 36 + 16)
renderDigit(draw, test_x[index][2], 0 * 28, index * 180 + 2 * 36)
renderImage(image, "MNIST_data/" + getOperator(result[2][:4]) + ".png",
2 * 28, index * 180 + 2 * 36)
renderActivations200(draw, activations_0[index][2], 3 * 28,
index * 180 + 2 * 36)
renderActivations200(draw, activations_1[index][2], 3 * 28,
index * 180 + 2 * 36 + 16)
renderDigit(draw, test_x[index][3], 0 * 28, index * 180 + 3 * 36)
renderImage(image,
"MNIST_data/" + str(argmax(result[3][:10])) + ".png",
1 * 28, index * 180 + 3 * 36)
renderImage(image,
"MNIST_data/" + str(argmax(result[3][10:20])) + ".png",
2 * 28, index * 180 + 3 * 36)
renderActivations200(draw, activations_0[index][3], 3 * 28,
index * 180 + 3 * 36)
renderActivations200(draw, activations_1[index][3], 3 * 28,
index * 180 + 3 * 36 + 16)
image.save(name + ".png")
def best_policy(mdp, U):
"""Dado um MDP e uma função de utilidade U, determinar a melhor política,
como um mapeamento do estado para a ação."""
pi = {}
for s in mdp.states:
pi[s] = argmax(mdp.actions(s),
key=lambda a: expected_utility(a, s, U, mdp))
return pi
def predict(example):
"""Predict the target value for example. Calculate probabilities for each
class and pick the max."""
def class_probability(targetval):
attr_dist = attr_dists[targetval]
return target_dist[targetval] * product(attr_dist[a] for a in example)
return argmax(target_dist.keys(), key=class_probability)
def predict(example):
"""Predict the target value for example. Calculate probabilities for each
class and pick the max."""
def class_probability(targetval):
attr_dist = attr_dists[targetval]
return target_dist[targetval] * product(attr_dist[a] for a in example)
return argmax(target_dist.keys(), key=class_probability)
def predict(example):
"""Predict the target value for example. Consider each possible value,
and pick the most likely by looking at each attribute independently."""
def class_probability(targetval):
return (target_dist[targetval] *
product(attr_dists[targetval, attr][example[attr]]
for attr in dataset.inputs))
return argmax(targetvals, key=class_probability)
def genetic_algorithm(self, problem, map_canvas):
""" Genetic Algorithm modified for the given problem """
def init_population(pop_number, gene_pool, state_length):
""" initialize population """
population = []
for i in range(pop_number):
population.append(utils.shuffled(gene_pool))
return population
def recombine(state_a, state_b):
""" recombine two problem states """
start = random.randint(0, len(state_a) - 1)
end = random.randint(start + 1, len(state_a))
new_state = state_a[start:end]
for city in state_b:
if city not in new_state:
new_state.append(city)
return new_state
def mutate(state, mutation_rate):
""" mutate problem states """
if random.uniform(0, 1) < mutation_rate:
sample = random.sample(range(len(state)), 2)
state[sample[0]], state[sample[1]] = state[sample[1]], state[sample[0]]
return state
def fitness_fn(state):
""" calculate fitness of a particular state """
fitness = problem.value(state)
return int((5600 + fitness) ** 2)
current = Node(problem.initial)
population = init_population(100, current.state, len(current.state))
all_time_best = current.state
while(1):
population = [mutate(recombine(*select(2, population, fitness_fn)), self.mutation_rate.get()) for i in range(len(population))]
current_best = utils.argmax(population, key=fitness_fn)
if fitness_fn(current_best) > fitness_fn(all_time_best):
all_time_best = current_best
self.cost.set("Cost = " + str('%0.3f' % (-1 * problem.value(all_time_best))))
map_canvas.delete('poly')
points = []
for city in current_best:
points.append(self.frame_locations[city][0])
points.append(self.frame_locations[city][1])
map_canvas.create_polygon(points, outline='red', width=1, fill='', tag='poly')
best_points = []
for city in all_time_best:
best_points.append(self.frame_locations[city][0])
best_points.append(self.frame_locations[city][1])
map_canvas.create_polygon(best_points, outline='red', width=3, fill='', tag='poly')
map_canvas.update()
map_canvas.after(self.speed.get())
def value_iteration(mdp, gamma, epsilon=0.001):
"""
Perform value iteration of Markov Descision Process MDP.
Parameters
----------
mpd : mdp.MPD object
Markov Descision Process
gamma : float > 0
Discount factor
epsilon : float > 0
Algorithm sensitivity.
Returns
-------
dict of state : float
Map from state to value, where state is in mdp.states.
"""
# -------- TASK 2.2 --------------------------------------------------------
# TASK: Implement the value iteration algorithm
# This corresponds to step 1 - 3 of the algorithm as described on the
# MyCourses pages
# The initial value for each state is 0.
# The return value, v, should be a dictionary mapping from states
# to values. This corresponds to v_{n+1} once the change is small
# enough in step 3 and the algorithm terminates.
#
# CODE HERE CODE HERE CODE HERE
#
termination = epsilon * (1 - gamma) / (2 * gamma)
v = {}
nv = {}
va = {}
n = 0
f = False
for s in mdp.states():
v[s] = 0
nv = v.copy()
while not f:
f = True
for s in mdp.states():
va = {
a: value_of(mdp, s, a, v, gamma)
for a in mdp.applicable_actions(s)
}
max = argmax(va)
nv[s] = va[max]
if all([((abs(nv[value] - v[value])) < termination) for value in v]):
return nv
v = nv.copy()
f = False
return v
def fitness_threshold(fitness_fn, f_thres, population):
if not f_thres:
return None
fittest_individual = argmax(population, key=fitness_fn)
if fitness_fn(fittest_individual) >= f_thres:
return fittest_individual
return None
def _policy(self, current_state): """Select an action epsilon-greedily.""" if random.uniform(0, 1) < self.eps: action = self.env.action_space.sample() else: action = argmax(range(self.env.action_space.n), lambda a : self.Q[self.current_state][a]) return action
def epsilon_greedy(self, q_values, epsilon):
sampled = random.uniform(0.0, 1.0)
action_set = list(self.env.get_action_set(
self.env.get_current_state()))
if sampled < epsilon:
return random.choice(action_set)
else:
return utils.argmax(
q_values[action_set]) # I'm breaking ties randomly
def _update(self, action, reward): """Update Q according to reward received.""" Q = self.Q maxaction = argmax(range(self.env.action_space.n), lambda a : Q[self.current_state][a] - Q[self.previous_state][action]) maxactiondiff = Q[self.current_state][maxaction] - Q[self.previous_state][action] Q[self.previous_state][action] += self.alpha*(self.A*reward + self.B*self.gamma*maxactiondiff)
def call(self,
inputs,
decoder_type='argmax',
k=25,
p=0.9,
temperature=0.8):
"""
Args:
inputs: it is a list of the previous token, memorized keys and values, and size of the previous tokens.
token shape = (1, 1, 2 (ID and position))
mem_k shape = (number of layers, 1, number of heads, attn hidden size, context size)
mem_v shape = (number of layers, 1, number of heads, context size, attn hidden size)
length = an integer
Returns:
next_token: shape = (1, 1, 2 (ID and position))
mem_k: shape = (number of layers, 1, number of heads, attn hidden size, context size)
mem_v: shape = (number of layers, 1, number of heads, context size, attn hidden size)
"""
token = inputs[0]
mem_k = inputs[1]
mem_v = inputs[2]
length = inputs[3]
new_mem_k = []
new_mem_v = []
hidden = self.embed(token)
for i, block in enumerate(self.transformer_stack):
hidden, k, v = block(hidden, mem_k[i], mem_v[i], length)
new_mem_k.append(k)
new_mem_v.append(v)
mem_k = tf.stack(new_mem_k)
mem_v = tf.stack(new_mem_v)
logit = tf.reshape(
tf.matmul(hidden[0, :, :],
self.embed.we[:self.n_vocab, :],
transpose_b=True), [self.n_vocab])
if decoder_type == 'argmax':
next_token = argmax(logit)
elif decoder_type == 'top-k':
next_token = top_k_sampling(logit, k, temperature)
elif decoder_type == 'nucleus':
next_token = nucleus_sampling(logit, p)
else:
next_token = sampling(logit, temperature)
return next_token, mem_k, mem_v
def genetic_algorithm(population,
fitness_fn,
gene_pool=[0, 1],
f_thres=None,
ngen=1000,
pmut=0.1):
"""[Figure 4.8]"""
for i in range(ngen):
population = [
mutate(recombine(*select(2, population, fitness_fn)), gene_pool,
pmut) for i in range(len(population))
]
if f_thres:
fittest_individual = argmax(population, key=fitness_fn)
if fitness_fn(fittest_individual) >= f_thres:
return fittest_individual
return argmax(population, key=fitness_fn)
def __init__(self, col, Y):
self.col = col
self.children = {}
self.prob = Counter(Y)
s = sum(self.prob.values())
for y in self.prob:
self.prob[y] /= s
label_ind, self.label_prob = argmax(self.prob.keys(),
key=self.prob.__getitem__)
self.label = Y[label_ind]
def predict(example):
"""Predict the target value for example. Consider each possible value,
and pick the most likely by looking at each attribute independently."""
def class_probability(targetval):
prob = target_dist[targetval]
for attr in dataset.inputs:
prob *= gaussian(means[targetval][attr], deviations[targetval][attr], example[attr])
return prob
return argmax(target_vals, key=class_probability)
def _viterbi_decode(self, feats):
backpointers = []
# Initialize the viterbi variables in log space
init_vvars = torch.Tensor(1, self.tagset_size).fill_(-10000.)
init_vvars[0][self.tag_to_ix[START_TAG]] = 0
# forward_var at step i holds the viterbi variables for step i-1
forward_var = autograd.Variable(init_vvars)
for feat in feats:
bptrs_t = [] # holds the backpointers for this step
viterbivars_t = [] # holds the viterbi variables for this step
for next_tag in range(self.tagset_size):
# next_tag_var[i] holds the viterbi variable for tag i at the
# previous step, plus the score of transitioning
# from tag i to next_tag.
# We don't include the emission scores here because the max
# does not depend on them (we add them in below)
next_tag_var = forward_var + self.transitions[next_tag]
best_tag_id = argmax(next_tag_var)
bptrs_t.append(best_tag_id)
viterbivars_t.append(next_tag_var[0][best_tag_id])
# Now add in the emission scores, and assign forward_var to the set
# of viterbi variables we just computed
forward_var = (torch.cat(viterbivars_t) + feat).view(1, -1)
backpointers.append(bptrs_t)
# Transition to STOP_TAG
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]
best_tag_id = argmax(terminal_var)
path_score = terminal_var[0][best_tag_id]
# Follow the back pointers to decode the best path.
best_path = [best_tag_id]
for bptrs_t in reversed(backpointers):
best_tag_id = bptrs_t[best_tag_id]
best_path.append(best_tag_id)
# Pop off the start tag (we dont want to return that to the caller)
start = best_path.pop()
assert start == self.tag_to_ix[START_TAG] # Sanity check
best_path.reverse()
return path_score, best_path
def predict(example):
"""Predict the target value for example. Consider each possible value,
and pick the most likely by looking at each attribute independently."""
def class_probability(targetval):
prob = target_dist[targetval]
for attr in dataset.inputs:
prob *= gaussian(means[targetval][attr], deviations[targetval][attr], example[attr])
return prob
return argmax(target_vals, key=class_probability)
def most_open_move(board, order):
"Find the move that has the most open floor filled around it."
p1, p2, t, touching = dfs_count_around(board)
wall_move = follow_wall_move(board, order)
open_move = utils.argmax(p1.keys(), lambda k: p1[k])
if p1[wall_move] == p1[open_move]:
best_move = wall_move
else:
best_move = open_move
logging.debug("most open move is: %s (%d) %s", best_move, p1[best_move], p1)
return best_move
def best_policy(mdp, U):
"""与えられたMDPと価値関数から最適な方策を決定する。
具体的には、全ての状態に関して{状態s:するべき行動a}の辞書を追加して返す。"""
pi = {}
for s in mdp.states:
# keyで比較を行なっている。つまり、sにおいて全ての(s,a)の価値を計算して比較。
# 最大のaを、{s:a}として登録。
pi[s] = argmax(mdp.actions(s),
key=lambda a: expected_utility(a, s, U, mdp))
return pi
def EpsilonGreedy(self, Q, s, epsilon, episode):
actions_in_state = self.actions_in_state
A = actions_in_state(s, episode)
a = argmax(A, key=lambda a1:Q[s,a1])
rand_num = random.uniform(0,1)
if rand_num < epsilon:
a = random.choice(A)
return a
def best_policy(mdp, U):
"""Given an MDP and a utility function U, determine the best policy,
as a mapping from state to action."""
pi = {}
f = lambda a: expected_utility(a, s, U, mdp)
for count, s in enumerate(mdp.states):
if count % 1 == 0:
#for s in mdp.states:
pi[s] = argmax(mdp.actions(s), key=f)
return pi
def execute(self, percept):
"""Execute the information gathering algorithm"""
self.observation = self.integrate_percept(percept)
vpis = self.vpi_cost_ratio(self.variables)
j = argmax(vpis)
variable = self.variables[j]
if self.vpi(variable) > self.cost(variable):
return self.request(variable)
return self.decnet.best_action()
def most_open_move(board, order):
"Find the move that has the most open floor filled around it."
p1, p2, t, touching = dfs_count_around(board)
wall_move = follow_wall_move(board, order)
open_move = utils.argmax(p1.keys(), lambda k: p1[k])
if p1[wall_move] == p1[open_move]:
best_move = wall_move
else:
best_move = open_move
logging.debug("most open move is: %s (%d) %s", best_move, p1[best_move],
p1)
return best_move
def __init__(self, col, Y):
self.col = col
self.children = {}
# Calculate the total
self.prob = Counter(Y)
s = sum(self.prob.values())
for y in self.prob:
# Standardization
self.prob[y] /= s
# calculate the largest
label_ind, self.label_prob = argmax(self.prob.keys(), key=self.prob.__getitem__)
self.labels = Y[label_ind]
def genetic_algorithm(population, fitness_fn, gene_pool=[0, 1], f_thres=None, ngen=1000, pmut=0.1): # noqa
"""[Figure 4.8]"""
for i in range(ngen):
new_population = []
random_selection = selection_chances(fitness_fn, population)
for j in range(len(population)):
x = random_selection()
y = random_selection()
child = reproduce(x, y)
if random.uniform(0, 1) < pmut:
child = mutate(child, gene_pool)
new_population.append(child)
population = new_population
if f_thres:
fittest_individual = argmax(population, key=fitness_fn)
if fitness_fn(fittest_individual) >= f_thres:
return fittest_individual
return argmax(population, key=fitness_fn)
def generate(self, in_str, max_len=2, **kwargs):
dy.renew_cg()
in_seq = u.preprocess_seq(in_str, self.char2int)
out_str, pred, EOSint = "", None, self.char2int[u.EOS]
decoder = self.make_decoder(in_seq, **kwargs)
probs = next(decoder)
while pred != EOSint and len(out_str) < (max_len * len(in_seq)):
pred = u.argmax(probs) # greedy search (take best prediction)
next(decoder)
probs = decoder.send(pred)
out_str += self.int2char[pred]
return out_str
def viterbi_decode(self, feats):
'''
In this function, we implement the viterbi algorithm explained above.
A Dynamic programming based approach to find the best tag sequence
'''
backpointers = []
# analogous to forward
# Initialize the viterbi variables in log space
init_vvars = torch.Tensor(1, self.tagset_size).fill_(-10000.)
init_vvars[0][self.tag_to_ix[START_TAG]] = 0
# forward_var at step i holds the viterbi variables for step i-1
forward_var = Variable(init_vvars)
if self.use_gpu:
forward_var = forward_var.cuda()
for feat in feats:
next_tag_var = forward_var.view(1, -1).expand(
self.tagset_size, self.tagset_size) + self.transitions
_, bptrs_t = torch.max(next_tag_var, dim=1)
bptrs_t = bptrs_t.squeeze().data.cpu().numpy(
) # holds the backpointers for this step
next_tag_var = next_tag_var.data.cpu().numpy()
viterbivars_t = next_tag_var[
range(len(bptrs_t)),
bptrs_t] # holds the viterbi variables for this step
viterbivars_t = Variable(torch.FloatTensor(viterbivars_t))
if self.use_gpu:
viterbivars_t = viterbivars_t.cuda()
# Now add in the emission scores, and assign forward_var to the set
# of viterbi variables we just computed
forward_var = viterbivars_t + feat
backpointers.append(bptrs_t)
# Transition to STOP_TAG
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]
terminal_var.data[self.tag_to_ix[STOP_TAG]] = -10000.
terminal_var.data[self.tag_to_ix[START_TAG]] = -10000.
best_tag_id = argmax(terminal_var.unsqueeze(0))
path_score = terminal_var[best_tag_id]
# Follow the back pointers to decode the best path.
best_path = [best_tag_id]
for bptrs_t in reversed(backpointers):
best_tag_id = bptrs_t[best_tag_id]
best_path.append(best_tag_id)
# Pop off the start tag (we dont want to return that to the caller)
start = best_path.pop()
assert start == self.tag_to_ix[START_TAG] # Sanity check
best_path.reverse()
return path_score, best_path
def play(self, game):
""" Returns the "best" move to play in the current <game>-state, after some deliberation (<check_abort>).
"""
def display(depth, label, parent_label, i):
for _ in range(depth):
print(' ', end='')
if parent_label is not None:
print(parent_label.moves[i],
parent_label.q[i].n,
parent_label.q[i].avg,
end=': ')
print(label.n)
return
def visits(q, i):
return q[i].n
self.reset()
if self.tree is None:
self.tree = tree.Tree()
in_advanced_mode = self._params.get('advanced')
max_num_simulations = self._params.get('simulations')
if max_num_simulations is None:
max_num_simulations = 0
num_simulations = 0
#while not check_abort.do_abort() and (max_num_simulations == 0 or num_simulations <= max_num_simulations):
while (max_num_simulations == 0
or num_simulations <= max_num_simulations):
mcts.simulate(self, game, self.tree, in_advanced_mode)
# tree.depth_first_traversal(self.tree, self.tree.root(), 0, display)
num_simulations += 1
node_id = self.tree.root()
node_label = self.tree.node_label(node_id)
max_i = utils.argmax(node_label.q, len(node_label.q), visits)
e = self._params.get('explore')
if e is not None and e > game.get_move_no() and random.randint(
1, 10) <= 8:
# Choose a random move with a 80% change for first e moves, if requested.
max_i = random.randint(0, node_label.len - 1)
policy = [node_label.q[i].n for i in range(node_label.len)]
total = 0
for p in policy:
total += p
if total > 0:
policy = [p / total for p in policy]
# tree.depth_first_traversal(self.tree, self.tree.root(), 0, display)
return node_label.moves[max_i], node_label.q[
max_i].avg, max_i, node_label.moves, policy, node_label.q
def _viterbi_decode(self, feats):
backpointers = []
# Initialize the viterbi variables in log space
init_vvars = torch.full((1, self.tagset_size), -10000.)
init_vvars[0][self.tag_to_ix[START_TAG]] = 0
# forward_var at step i holds the viterbi variables for step i-1
forward_var = init_vvars
for feat in feats:
bptrs_t = [] # holds the backpointers for this step
viterbivars_t = [] # holds the viterbi variables for this step
for next_tag in range(self.tagset_size):
# next_tag_var[i] holds the viterbi variable for tag i at the
# previous step, plus the score of transitioning
# from tag i to next_tag.
next_tag_var = forward_var + self.transitions[next_tag]
best_tag_id = argmax(next_tag_var)
bptrs_t.append(best_tag_id)
viterbivars_t.append(next_tag_var[0][best_tag_id].view(1))
forward_var = (torch.cat(viterbivars_t) + feat).view(1, -1)
backpointers.append(bptrs_t)
# Transition to STOP_TAG
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]
best_tag_id = argmax(terminal_var)
path_score = terminal_var[0][best_tag_id]
# Follow the back pointers to decode the best path.
best_path = [best_tag_id]
for bptrs_t in reversed(backpointers):
best_tag_id = bptrs_t[best_tag_id]
best_path.append(best_tag_id)
# Pop off the start tag (we dont want to return that to the caller)
start = best_path.pop()
assert start == self.tag_to_ix[START_TAG] # Sanity check
best_path.reverse()
return path_score, best_path
def genetic_algorithm(population, fitness_fn, ngen=1000, pmut=0.1):
"[Fig. 4.8]"
for i in range(ngen):
new_population = []
for i in len(population):
fitnesses = map(fitness_fn, population)
p1, p2 = weighted_sample_with_replacement(population, fitnesses, 2)
child = p1.mate(p2)
if random.uniform(0, 1) < pmut:
child.mutate()
new_population.append(child)
population = new_population
return argmax(population, fitness_fn)
def sample_site_cftp_dep(matrix, mu, Ne):
L = len(matrix)
def log_phat(s):
ep = score_seq(matrix,s)
nu = Ne - 1
return -nu*log(1 + exp(ep - mu))
first_site = "A"*L
last_site = "T"*L
best_site = "".join(["ACGT"[argmin(row)] for row in matrix])
worst_site = "".join(["ACGT"[argmax(row)] for row in matrix])
trajs = [[best_site],[random_site(L)],[random_site(L)],[random_site(L)], [worst_site]]
def mutate_site(site,(ri,rb)):
return subst(site,"ACGT"[rb],ri)
def genetic_algorithm(population, fitness_fn, ngen=1000, pmut=0.1):
"[Figure 4.8]"
for i in range(ngen):
new_population = []
for i in range(len(population)):
fitnesses = map(fitness_fn, population)
p1, p2 = weighted_sample_with_replacement(population, fitnesses, 2)
child = p1.mate(p2)
if random.uniform(0, 1) < pmut:
child.mutate()
new_population.append(child)
population = new_population
return argmax(population, key=fitness_fn)
def policy_iteration(mdp):
"Solve an MDP by policy iteration [Figure 17.7]"
U = {s: 0 for s in mdp.states}
pi = {s: random.choice(mdp.actions(s)) for s in mdp.states}
while True:
U = policy_evaluation(pi, U, mdp)
unchanged = True
for s in mdp.states:
a = argmax(mdp.actions(s), key=lambda a: expected_utility(a, s, U, mdp))
if a != pi[s]:
pi[s] = a
unchanged = False
if unchanged:
return pi
def max_value(node):
if game.terminal_test(node):
return game.utility(node, player)
self.change_list.append(('a', node))
self.change_list.append(('h',))
max_a = argmax(game.actions(node), key=lambda x: min_value(game.result(node, x)))
max_node = game.result(node, max_a)
self.utils[node] = self.utils[max_node]
x1, y1 = self.node_pos[node]
x2, y2 = self.node_pos[max_node]
self.change_list.append(('l', (node, max_node - 3*node - 1)))
self.change_list.append(('e', node))
self.change_list.append(('p',))
self.change_list.append(('h',))
return self.utils[node]
def program(self,percept):
s1,r1 = percept
Q, Nsa, s, a, r = self.Q, self.Nsa, self.s, self.a, self.r
alpha, gamma, f = self.alpha, self.gamma, self.f
if s1 in self.terminals:
Q[s1][None] = r1
if s is not None:
Nsa[s][a] += 1
Q[s][a] += alpha(Nsa[s][a])*(r+gamma*max(Q[s1].values())-Q[s][a])
if s1 in self.terminals:
self.s = self.a = self.r = None
else:
self.s, self.r = s1, r1
self.a = argmax(Q[s1].keys(), lambda a1: f(Q[s1][a1],Nsa[s1][a1]))
return self.a
def __call__(self, percept):
s1, r1 = self.update_state(percept)
Q, Nsa, s, a, r = self.Q, self.Nsa, self.s, self.a, self.r
alpha, gamma, terminals, actions_in_state = self.alpha, self.gamma, self.terminals, self.actions_in_state
if s1 in terminals:
Q[(s1, None)] = r1
if s is not None:
Nsa[(s, a)] += 1
Q[(s, a)] += alpha(Nsa[(s, a)])*(r+gamma*max([Q[(s1, a1)] for a1 in actions_in_state(s1)])-Q[(s, a)])
if s1 in terminals:
self.s = self.a = self.r = None
else:
self.s, self.r = s1, r1
self.a = argmax(actions_in_state(s1), key=lambda a1: self.f(Q[(s1, a1)], Nsa[(s1, a1)]))
return self.a
def _init_clusters(self): sub_docs = self.docs[:self.num_instances] sim_mat = utils.pairwise(sub_docs, lambda x, y: max(self.doc_similarity(x, y), self.doc_similarity(y, x))) edges = utils.minimum_spanning_tree(sim_mat) ccs = utils.get_ccs(range(self.num_instances), edges) biggest_cc = max(map(len, ccs)) while biggest_cc > self.num_init: edge_to_remove = random.sample(edges, 1)[0] edges.remove(edge_to_remove) ccs = utils.get_ccs(range(self.num_instances), edges) biggest_cc = max(map(len, ccs)) cc = ccs[utils.argmax(map(len, ccs))] for idx in cc: self._add_cluster(self.docs[idx], member=False)
def sample_site_spec(matrix, mu, Ne):
nu = Ne - 1
L = len(matrix)
best_site = "".join(["ACGT"[argmin(col)] for col in matrix])
worst_site = "".join(["ACGT"[argmax(col)] for col in matrix])
def phat(s):
assert len(s) == L
ep = score_seq(matrix,s)
return (1 + exp(ep - mu))**(-nu)
chosen_site = ""
def best_completion(s):
l = len(s)
return phat(s + best_site[l:])
def worst_completion(s):
l = len(s)
return s + worst_site[l:]
return chosen_site
def sample_site_cftp(matrix, mu, Ne):
L = len(matrix)
f = seq_scorer(matrix)
def log_phat(s):
ep = f(s)
nu = Ne - 1
return -nu*log(1 + exp(ep - mu))
first_site = "A"*L
last_site = "T"*L
best_site = "".join(["ACGT"[argmin(row)] for row in matrix])
worst_site = "".join(["ACGT"[argmax(row)] for row in matrix])
#middle_sites = [[random_site(L)] for i in range(10)]
#trajs = [[best_site]] + middle_sites + [[worst_site]]
trajs = [[best_site],[worst_site]]
ords = [rslice("ACGT",sorted_indices(row)) for row in matrix]
def mutate_site(site,(ri,direction)):
b = (site[ri])
idx = ords[ri].index(b)
idxp = min(max(idx + direction,0),3)
bp = ords[ri][idxp]
return subst(site,bp,ri)
def find_mean_field_approximation(ks,q,ps=None):
eps = eps_from_ks(ks)
if ps is None:
ps = occupancies_ref(ks,q)
mu_min = -10
mu_max = 10
mu_steps = 1000
mus = interpolate(mu_min,mu_max,mu_steps)
coeffs = map(lambda mu:polyfit(ps,probs(eps,mu),1)[0],
mus)
max_coeff_idx = argmax(coeffs)
# find mu corresponding to best fit. Cutoff at peak, since curve
# has two intersections with y = 1.
mu_idx = argmin(map(lambda coeff:(1-coeff)**2,coeffs[:max_coeff_idx]))
best_mu = mus[mu_idx]
qs = probs(eps,best_mu)
print "best_mu:",best_mu
print "mean copy number: %s (sd: %s) vs. sum(ps) %s" %(mean_occ(eps,best_mu),sd_occ(eps,best_mu),sum(ps))
print "pearson correlation:",pearsonr(ps,qs)
print "best linear fit: p = %s*q + %s" % tuple(polyfit(qs,ps,1))
return best_mu
def expectiminimax(state, game):
"""Return the best move for a player after dice are thrown. The game tree
includes chance nodes along with min and max nodes. [Figure 5.11]"""
player = game.to_move(state)
def max_value(state, dice_roll):
v = -infinity
for a in game.actions(state):
v = max(v, chance_node(state, a))
game.dice_roll = dice_roll
return v
def min_value(state, dice_roll):
v = infinity
for a in game.actions(state):
v = min(v, chance_node(state, a))
game.dice_roll = dice_roll
return v
def chance_node(state, action):
res_state = game.result(state, action)
if game.terminal_test(res_state):
return game.utility(res_state, player)
sum_chances = 0
num_chances = 21
dice_rolls = list(itertools.combinations_with_replacement([1, 2, 3, 4, 5, 6], 2))
if res_state.to_move == 'W':
for val in dice_rolls:
game.dice_roll = (-val[0], -val[1])
sum_chances += max_value(res_state,
(-val[0], -val[1])) * (1/36 if val[0] == val[1] else 1/18)
elif res_state.to_move == 'B':
for val in dice_rolls:
game.dice_roll = val
sum_chances += min_value(res_state, val) * (1/36 if val[0] == val[1] else 1/18)
return sum_chances / num_chances
# Body of expectiminimax:
return argmax(game.actions(state),
key=lambda a: chance_node(state, a))
