def reproduce(self):
"""
This fun creates N new nodes and assigns regions (i.e. hypercubes) to them.
:return: A list of the N new nodes.
"""
if len(self.hypercube_list) == 1:
new_hypercubes = []
new_hypercube_length = self.hypercube_list[0].length / 2
old_center = self.hypercube_list[0].center
num_new_hypercubes = 2**self.dimension
for i in range(num_new_hypercubes):
center_translation = np.fromiter(map(
lambda x: new_hypercube_length / 2
if x == '1' else -new_hypercube_length / 2,
list(bin(i)[2:].zfill(self.dimension))),
dtype=np.float)
new_hypercubes.append(
Hypercube(new_hypercube_length,
old_center + center_translation))
return [
UcbNode(self, self.h + 1,
new_hypercubes[:int(num_new_hypercubes / 2)]),
UcbNode(self, self.h + 1,
new_hypercubes[int(num_new_hypercubes / 2):])
]
else:
return [
UcbNode(
self, self.h + 1,
self.hypercube_list[:int(len(self.hypercube_list) / 2)]),
UcbNode(
self, self.h + 1,
self.hypercube_list[int(len(self.hypercube_list) / 2):])
]
def calculate_m(self, current_segment):
results = dict()
for neighbor in self.hypercube.backward_neighbors(current_segment):
# calculate M + S for each neighbor
M = neighbor.get_distance()
indices_changed = Hypercube.compare_indices(current_segment.index, neighbor.index)
sequence = list()
for i in range(0, len(indices_changed)):
if indices_changed[i]:
sequence.append(self.hypercube.sequences[i][current_segment.index[i]])
else:
sequence.append('-')
result = (M + self.calculate_s(sequence), neighbor.index)
if result[0] not in results:
results[result[0]] = list()
results[result[0]].append(result[1])
else:
results[result[0]].append(result[1])
if len(results) <= 0:
return None
max_result = max(results.keys())
backtrack_index = results[max_result]
return (max_result, backtrack_index)
def generate_output(self, traceback_path):
output_sequences = list()
while len(output_sequences) < self.hypercube.dimension_count:
output_sequences.append(list())
sequence_indices = list(self.hypercube.dimensions)
previous_step = traceback_path[0]
for current_step in traceback_path[1:]:
indices_changed = Hypercube.compare_indices(previous_step, current_step)
previous_step = tuple(current_step)
for i in range(0, len(indices_changed)):
if indices_changed[i]:
sequence_indices[i] -= 1
if sequence_indices[i] >= 0:
output_sequences[i].insert(0, self.hypercube.sequences[i][sequence_indices[i]])
else:
output_sequences[i].insert(0, '-')
else:
output_sequences[i].insert(0, '-')
return output_sequences
use_saved_data = True # when True, the script simply plots the data of the most recently ran simulation, if available
# this means that no simulations are run when True.
available_arms_mean = 50
num_times_to_run = 10
num_rounds = 50000
num_std_to_show = 5
budgets = [2, 4]
line_style_dict = {2: '-', 4: ':'}
v1 = np.sqrt(5)
v2 = 1
rho = 0.5
N = 2 # changing this HAS NO EFFECT and the ACC-UCB class will not work when N != 2
root_context = Hypercube(1,
np.array([0.5, 0.5
])) # this is called x_{0,1} in the paper
def run_one_try(problem_model, num_run, budget):
random_algo = Random(problem_model, budget)
bench_algo = Benchmark(problem_model, budget)
cc_mab_algo = CCMAB(problem_model, budget, root_context.get_dimension())
cc_ucb_algo = ACCUCB(problem_model, v1, v2, N, rho, budget, root_context)
ucb_reward, ucb_regret = cc_ucb_algo.run_algorithm()
bench_reward, bench_regret = bench_algo.run_algorithm()
random_reward, random_regret = random_algo.run_algorithm()
mab_reward, mab_regret = cc_mab_algo.run_algorithm()
print("Run done: " + str(num_run))
