def plan_tour(xys, budget, exact, fixed=[]):
"""
Calculates a tour over all given points.
:param xys: list of points to visit
:param budget: limit of points to visit + length limit
:param exact: whether or not to run Gurobi
:param fixed: segments of the trip already travelled (using this option avoids heuristic mode [multistart_localsearch])
:return: solution, feasibility, cost of the tour including probings, whether the tour is optimal
"""
n = len(xys)
pos = xys
c = {}
for i in range(n):
for j in range(n): # non optimized distance calculations!
c[i, j] = dist(*pos[i], *pos[j])
c[j, i] = c[i, j]
cost_is_optimal_or_timeout = False
if n > 2:
# heuristic
if len(fixed) == 0:
sol_, cost = multistart_localsearch(100, n, c, cutoff=budget -
n) # inst.T - (N) * inst.t)
idx = sol_.index(0)
sol = sol_[idx:] + sol_[:idx]
cost = complete_cost(cost, n)
# exact (or time out)
if len(fixed) > 0 or cost > budget:
if exact:
print("# trying exact solution")
cost, edges = solve_tsp(
range(n), c, False, fixed,
60000) # not using cutoff because it appears to be slow
# cost, edges = tsp(n, c, cutoff=budget - n)
cost = complete_cost(cost, n)
cost_is_optimal_or_timeout = True
try:
sol = sequence(range(n), edges)
except ValueError:
print("# time out")
cost = 99999999
sol = []
# keeps cost_is_optimal_or_timeout=True to avoid adding more points in add_while_possible()
else:
pass
print('# NOT trying exact solution')
elif n == 1:
cost = 0
sol = [0]
cost_is_optimal_or_timeout = True
else:
cost = dist(*pos[0], *pos[1])
cost = complete_cost(cost, n)
sol = [0, 1]
cost_is_optimal_or_timeout = True
return sol, cost <= budget, cost, cost_is_optimal_or_timeout
# random construction
t = randtour(n) # create a random tour
z = length(t,D) # calculate its length
update_graph(z,t,"random solution")
print("random:", t, z, ' --> ', end=' ')
z = localsearch(t,z,D) # local search starting from the random tour
print(t, z)
update_graph(z,t,"random solution, after local search")
# greedy construction
# [try cycle for more possibilities] for i in range(n):
i = 0
t = nearest_neighbor(n,i,D) # create a greedy tour, visiting city 'i' first
z = length(t,D)
update_graph(z,t,"nearest_neighbor, starting on city %d" % i)
print("nneigh:", t, z, ' --> ', end=' ')
z = localsearch(t,z,D)
print(t, z)
update_graph(z,t,"nearest_neighbor, starting on city %d (after local search)" % i)
# multi-start local search
print("random start local search:")
niter = 1000
t,z = multistart_localsearch(niter,n,D,update_graph)
print("best found solution on %d multistart_localsearch:" % niter)
print(t)
print("z=",z)
update_graph(z,t,"best found solution on %d multistart_localsearch:" % niter)
input("[press enter for exiting]")
def planner(X, z, f, dynam=False):
"""planner: decide list of points to visit based on:
- X: list of coordinates [(x1,y1), ...., (xN,yN)]
- z: list of evaluations of "true" function [z1, ..., zN]
- f: useless in static version
"""
X = list(X) # work with local copies
z = list(z)
from tsp import solve_tsp, sequence # exact
from tsp import multistart_localsearch # heuristic
kernel = RationalQuadratic(length_scale_bounds=(0.08, 100)) + WhiteKernel(
noise_level_bounds=(1e-5, 1e-2))
gpr = GaussianProcessRegressor(kernel=kernel, n_restarts_optimizer=10)
# # plot preliminary GP
# from functions import plot
# from functions import f1
# plot(f1,100)
# # end of plot
# # # plot posteriori GP
# GPR = GaussianProcessRegressor(kernel=kernel, n_restarts_optimizer=10)
# GPR.fit(X, z)
# def GP(x_,y_):
# return GPR.predict([(x_,y_)])[0]
# plot(GP,100)
# # # end of plot
x0, y0 = (inst.x0, inst.y0) # initial position (depot)
pos = [(x0, y0)]
N = 1
c = {}
while True:
# attempt adding Nth "city" to the tour
# retorna lista ordenada por variância
points = max_var(gpr, n=100) # n x n grid !!!!!
# pega ponto com maior variância
(x, y), (z_new_std, z_new) = points.pop()
# descarta previamente já selecionados, passando para a próxima maior variância
while (x, y) in pos:
(x, y), (z_new_std, z_new) = points.pop()
# estende matriz de distâncias para ambos solvers
for i in range(N):
c[i, N] = dist(*pos[i], x, y)
c[N, i] = c[i, N]
# evita TSP em menos de 3 'cidades', falta checar limite de tempo
if N < 3:
N += 1
pos.append((x, y))
continue # !!!!!
sol_, cost = multistart_localsearch(100,
N + 1,
c,
cutoff=inst.T -
(N) * inst.t) # heuristic
if cost <= inst.T - (N) * inst.t:
print("heuristic solution")
idx = sol_.index(0)
sol = sol_[idx:] + sol_[:idx] # heuristic
# print(obj + (N)*inst.t, "TSP solution:", obj, N, inst.T, sol)
# print("appending", (x,y), z_new_std, z_new, "orient.len:", obj + (N)*inst.t)
N += 1
assert (x, y) not in pos
pos.append((x, y))
X.append((x, y))
if (dynam):
z.append(f(x, y)) # !!!!! with PROBING
else:
z.append(
z_new
) # !!!!! with average of the prediction as an estimate
gpr.fit(X, z)
else:
# attempt exact solution:
print("attempting exact solution")
cost, edges = solve_tsp(range(N + 1), c) # exact
if cost <= inst.T - (N) * inst.t:
sol = sequence(range(N + 1), edges) # exact
# print(obj + (N) * inst.t, "TSP EXACT:", obj, N, inst.T, sol)
# print("appending", (x, y), z_new_std, z_new, "orient.len:", obj + (N) * inst.t)
N += 1
pos.append((x, y))
X.append((x, y))
if (dynam):
z.append(f(x, y)) # !!!!! with PROBING
else:
z.append(
z_new
) # !!!!! with average of the prediction as an estimate
gpr.fit(X, z)
print("found; continue")
continue
print("heuristic and exact solution exceeds time limit")
# print("testing next interesting points")
print("break")
break
print(cost + (N) * inst.t, "TSP solution:", cost, N, inst.T, sol)
return [pos[i] for i in sol[1:]]
def planner(X, z, f, mode, dynam=False):
"""planner: decide list of points to visit based on:
- X: list of coordinates [(x1,y1), ...., (xN,yN)]
- z: list of evaluations of "true" function [z1, ..., zN]
- f: useless in static version
"""
X = list(X) # work with local copies
z = list(z)
l = len(X)
kernel = RationalQuadratic(length_scale_bounds=(0.08, 100)) + WhiteKernel(
noise_level_bounds=(1e-5, 1e-2))
gpr = GaussianProcessRegressor(kernel=kernel, n_restarts_optimizer=10)
#gpr.fit(X,z) <- we should be fitting but for some reason it often worsens the solution
# # plot preliminary GP
# from functions import plot
# from functions import f1
# plot(f1,100)
# # end of plot
# # # plot posteriori GP
# GPR = GaussianProcessRegressor(kernel=kernel, n_restarts_optimizer=10)
# GPR.fit(X, z)
# def GP(x_,y_):
# return GPR.predict([(x_,y_)])[0]
# plot(GP,100)
# # # end of plot
x0, y0 = (inst.x0, inst.y0) # initial position (depot)
pos = [(x0, y0)]
N = 1
c = {}
randomize = False if mode == 0 else True
local_search = False if mode % 2 == 0 else True
multimodal = False if mode < 2 else True
while True:
# retorna lista ordenada por variância
points = max_var(gpr, n=100) # n x n grid !!!!!
z_std = [i[1][0] for i in points]
if randomize:
if not multimodal:
(x, y), (z_new_std, z_new) = points.pop()
while (x, y) in pos:
(x, y), (z_new_std, z_new) = points.pop()
(x, y), (z_new_std, z_new) = search_around_point(x,
y,
pos,
gpr,
n=100)
else:
(x, y), (z_new_std, z_new) = get_next_point(points,
pos,
gpr,
local_search,
n=100)
#(x,y), (z_new_std, z_new) = find_next_candidate_point(pos, gpr, 0.01, 2)
else:
# pega ponto com maior variância
(x, y), (z_new_std, z_new) = points.pop()
# descarta previamente já selecionados, passando para a próxima maior variância
while (x, y) in pos:
(x, y), (z_new_std, z_new) = points.pop()
print('Trying to probe point nº ', N)
# estende matriz de distâncias para ambos solvers
for i in range(N):
c[i, N] = dist(
*pos[i], x, y
) / inst.s # We divide by the speed to reflect the time needed to travel
c[N, i] = c[i, N]
# evita TSP em menos de 3 'cidades', falta checar limite de tempo
if N < 3:
N += 1
pos.append((x, y))
X.append((x, y))
z.append(z_new)
#gpr.fit(X,z)
continue # !!!!!
sol_, cost = multistart_localsearch(100,
N + 1,
c,
cutoff=inst.T -
(N) * inst.t) # heuristic
if cost <= inst.T - (N) * inst.t:
print("heuristic solution")
idx = sol_.index(0)
sol = sol_[idx:] + sol_[:idx] # heuristic
# print(obj + (N)*inst.t, "TSP solution:", obj, N, inst.T, sol)
# print("appending", (x,y), z_new_std, z_new, "orient.len:", obj + (N)*inst.t)
N += 1
assert (x, y) not in pos
pos.append((x, y))
X.append((x, y))
if (dynam):
z.append(f(x, y)) # !!!!! with PROBING
else:
z.append(
z_new
) # !!!!! with average of the prediction as an estimate
gpr.fit(X, z)
else:
# attempt exact solution:
print("attempting exact solution")
cost, edges = solve_tsp(range(N + 1), c) # exact
if cost <= inst.T - (N) * inst.t:
sol = sequence(range(N + 1), edges) # exact
# print(obj + (N) * inst.t, "TSP EXACT:", obj, N, inst.T, sol)
# print("appending", (x, y), z_new_std, z_new, "orient.len:", obj + (N) * inst.t)
N += 1
pos.append((x, y))
X.append((x, y))
if (dynam):
z.append(f(x, y)) # !!!!! with PROBING
else:
z.append(
z_new
) # !!!!! with average of the prediction as an estimate
gpr.fit(X, z)
print("found; continue")
continue
print("heuristic and exact solution exceeds time limit")
# print("testing next interesting points")
print("break")
break
#print("Found original points. Checking them")
#print()
#result = check_chosen_points(X, pos, l, gpr, z, c, sol, N, n=100)
#pos, sol = result[0], result[1]
print(cost + (N) * inst.t, "TSP solution:", cost, N, inst.T, sol)
return [pos[i] for i in sol[1:]]