def __init__(self, max_size, width, height, num_channels, alpha):
self.max_size = max_size
#frames/images are stored in sequence, i.e. frames[i+1] is the state obtained by executing action actions[i+1] in state frames[i]
self.frames = np.zeros([max_size, width, height, num_channels],
dtype=np.float32)
self.actions = np.zeros(max_size, dtype=int)
self.rewards = np.zeros([max_size, 1], dtype=np.float32)
self.terminal = np.zeros([max_size, 1], dtype=int)
self.size = 0
self.next_index = 0
#use the next power of 2 as the size for the sum tree/max heap, this simplifies their implementation
power_2_size = self.get_next_power(max_size)
self.sum_tree = SumTree(power_2_size, alpha)
self.max_heap = MaxHeap(power_2_size)
#additive constant to keep td-values above 0
self.td_epsilon = 1e-9
import time
import numpy
import sys
budget = 20
graph_type = int(sys.argv[1])
stats = [[] for _ in range(budget)]
ans = [[] for _ in range(budget)]
for i in range(1, 51):
filename = "../DATA/mass_data/input" + `graph_type` + "-" + `i` + ".txt"
G = read_file(filename)
start_time = time.time()
G.preprocess()
PQ = MaxHeap()
for i in range(G.N):
PQ.update(i, G.N * G.N)
S = set()
cur_val = G.N * G.N
used = set()
for curr_budget in range(budget):
while True:
cur, val, aux = PQ.pop()
if cur == -1: break
if cur in used: continue
new_val = G.avg_dist(S.union(set([cur])))
class ExperienceMemory:
def __init__(self, max_size, width, height, num_channels, alpha):
self.max_size = max_size
#frames/images are stored in sequence, i.e. frames[i+1] is the state obtained by executing action actions[i+1] in state frames[i]
self.frames = np.zeros([max_size, width, height, num_channels],
dtype=np.float32)
self.actions = np.zeros(max_size, dtype=int)
self.rewards = np.zeros([max_size, 1], dtype=np.float32)
self.terminal = np.zeros([max_size, 1], dtype=int)
self.size = 0
self.next_index = 0
#use the next power of 2 as the size for the sum tree/max heap, this simplifies their implementation
power_2_size = self.get_next_power(max_size)
self.sum_tree = SumTree(power_2_size, alpha)
self.max_heap = MaxHeap(power_2_size)
#additive constant to keep td-values above 0
self.td_epsilon = 1e-9
def add(self, s, a, r, t):
i = self.next_index
self.frames[i, :, :] = s
self.actions[i] = a
self.rewards[i] = r
self.terminal[i] = t
#a new element gets maximum priority value
#this ensures that each transition is replayed at least once (with high probability)
p_value = 1
if self.size > 0:
p_value = self.max_heap.get_max()
self.sum_tree.update(i, p_value)
self.max_heap.update(i, p_value)
if self.size < self.max_size:
self.size += 1
self.next_index = (self.next_index + 1) % self.max_size
def sample(self, n):
#if the replay memory is not full we should not sample the frame 0 since it has no predecessor frame
min_index = 1
max_index = self.size
if self.size == self.max_size:
min_index = 0
i, p_values = self.sum_tree.sample(n)
i_minus = (i - 1) % self.size
s = self.frames[i_minus, :, :, :]
a = self.actions[i]
r = self.rewards[i]
s2 = self.frames[i, :, :, :]
t = self.terminal[i]
return s, a, r, s2, t, i, p_values
#update priority value of replayed transitions with their new td-error
def update_p(self, indices, td):
for i in range(len(indices)):
self.sum_tree.update(indices[i], td[i] + self.td_epsilon)
self.max_heap.update(indices[i], td[i] + self.td_epsilon)
#returns y >= x such that y is a power of 2
def get_next_power(self, x):
result = 1
if x > 0:
x_int = int(x - 1)
while x_int > 0:
x_int = x_int >> 1
result = result << 1
return result
from util import (
read_file,
MaxHeap
)
G = read_file("../OLD/input.txt")
G.preprocess()
budget = 10
seed = set()
PQ = MaxHeap()
last_seed = None
cur_best = None
cur_val = G.N * G.N
INIT_VAL = G.N * G.N
# initialize
for i in range(G.N):
mg1 = G.avg_dist(set([i]))
prev_best = cur_best
if cur_best is not None:
mg2 = G.avg_dist(set([i]).union(set([cur_best])))
else:
mg2 = mg1
flag = 0
PQ.update((i, mg1, prev_best, mg2, flag), mg1)
if mg1 < cur_val:
cur_val = mg1
cur_best = i
# update
while len(seed) < budget:
