def min_with_nn(f, input_box, esp, func):
game = BB(f, input_box, Interval(-999, 999), func)
nnet = nn(game)
nnet.load_checkpoint('./', 'best.pth.tar')
current_box = input_box
board = game.getBoardFromInput_box(current_box)
r = game.getGameEnded(current_box, esp)
step = 0
while r == 0:
board = game.getBoardFromInput_box(current_box)
#print(board)
pi, v = nnet.predict(board)
pi = game.getValidMoves(current_box, esp) * pi
#print(pi)
a = np.argmax(pi)
#print(a)
current_box = game.getNextState(current_box, a)
r = game.getGameEnded(current_box, esp)
step += 1
return r, step
def bnb_with_nn_pq(f, input_box, eps, func):
QSA = {}
game = BB(f, input_box, Interval(-999, 999), func)
nnet = nn(game)
nnet.load_checkpoint('./', 'best.pth.tar')
current_box = input_box
board = game.getBoardFromInput_box(current_box)
pi, v = nnet.predict(board)
box_stack = [(v, interval(current_box))]
QSA[game.stringRepresentation(current_box)] = (pi, v)
g_min = eval_at_mid(f, current_box)
step = 0
while box_stack:
v, cur_interval = heappop(box_stack)
current_box = cur_interval.interval
if g_min < get_lb(f, current_box):
continue
s = game.stringRepresentation(current_box)
pi, v = QSA[s]
pi = game.getValidMoves(current_box, eps) * pi
if np.sum(pi) > 0:
a = np.argmax(pi)
if pi[a] > 0:
pi[a] = 0
QSA[s] = (pi, v)
if a % 2 == 0:
alter_a = a + 1
else:
alter_a = a - 1
alter_box = game.getNextState(current_box, alter_a)
current_box = game.getNextState(current_box, a)
new_boxes = [alter_box, current_box]
for b in new_boxes:
if eval_at_mid(f, b) < g_min - eps:
g_min = eval_at_mid(f, b)
#only push the box if its overestimation of lower bound is lower
if get_lb(f, b) < g_min - eps:
s = game.stringRepresentation(b)
if s not in QSA:
board = game.getBoardFromInput_box(b)
pi, v = nnet.predict(board)
QSA[s] = (pi, v)
if b not in box_stack:
heappush(box_stack, (v, interval(b)))
step += 1
return g_min, step
#-20 * exp(-0.2*sqrt(0.5*(x^2+y^2)))-exp(0.5*(cos(2*3.1415926535*x)+cos(2*3.1415926535*y)))+2.71828+20"
i = 0
generator = GF(["x1", "x2"], [3, 3], -5, 5, 20)
generator.randomPara()
function = generator.generateString(generator.coe, generator.degree_matrix)
f = Function("x1", "x2", function)
#f = Function("x", "y", "-20 * exp(-0.2*sqrt(0.5*(x^2+y^2)))-exp(0.5*(cos(2*3.1415926535*x)+cos(2*3.1415926535*y)))+2.71828+20")
#Define the input domain of the function -- both[0.5,5] for x and y
input_box = IntervalVector(2, [-5, 5])
#Define the output range (i.e. desired value of the function) -- f range [1,1]
output_range = Interval(-999, 999)
g = BB(f, input_box, output_range, generator.generateFunc)
nnet = nn(g)
if args.load_model:
nnet.load_checkpoint(args.load_folder_file[0],
args.load_folder_file[1])
c = Coach(g, nnet, args, i)
i += 1
c.learn()
while True:
generator = GF(["x1", "x2"], [3, 3], -5, 5, 20)
generator.randomPara()
function = generator.generateString(generator.coe,
from naive.BB import BB
from naive.pytorch.NNet import NNetWrapper as nn
from pyibex import *
THRESHOLD = 0.001
f = Function("x", "y", "-20 * exp(-0.2*sqrt(0.5*(x^2+y^2)))-exp(0.5*(cos(2*3.1415926535*x)+cos(2*3.1415926535*y)))+2.71828+20")
#Define the input domain of the function -- both[0.5,5] for x and y
input_box = IntervalVector([[-5,5],[-4,3]])
#Define the output range (i.e. desired value of the function) -- f range [1,1]
output_range = Interval(0,10)
game = BB(f, input_box, output_range)
nnet = nn(game)
nnet.load_checkpoint('temp/', 'best.pth.tar')
current_box = input_box
board = game.getBoardFromInput_box(current_box)
r = game.getGameEnded(current_box, THRESHOLD)
while r == 0:
board = game.getBoardFromInput_box(current_box)
pi, v = nnet.predict(board)
pi = game.getValidMoves(current_box, THRESHOLD) * pi
print(pi)
a = np.argmax(pi)
current_box = game.getNextState(current_box, a)