def get_overview_mutation(of):
overview_mutations = []
for correction in [Correction(of), MirrorCorrection(of)]:
for r in [0.1, 0.5, 0.75]:
overview_mutations.append(
CauchyMutation(r, correction=Correction(of)))
return overview_mutations
def get_correction_mutation(of):
correction_mutations = []
for correction in [Correction(of), MirrorCorrection(of)]:
for r in [1, 1.5, 2, 5]:
correction_mutations.append(
CauchyMutation(r, correction=Correction(of)))
return correction_mutations
def experiment_fsa(of, maxeval, num_runs, T0, n0, alpha, r):
results = []
for i in tqdm(
range(num_runs),
'Testing T0={}, n0={}, alpha={}, r={}'.format(T0, n0, alpha, r)):
mut = CauchyMutation(r=r, correction=Correction(of))
result = FastSimulatedAnnealing(of,
maxeval=maxeval,
T0=T0,
n0=n0,
alpha=alpha,
mutation=mut).search()
result['run'] = i
result['heur'] = 'FSA_{}_{}_{}_{}'.format(T0, n0, alpha,
r) # name of the heuristic
result['T0'] = T0
result['n0'] = n0
result['alpha'] = alpha
result['r'] = r
results.append(result)
return pd.DataFrame(results,
columns=[
'heur', 'run', 'T0', 'n0', 'alpha', 'r', 'best_x',
'best_y', 'neval'
])
#
# In the current settings there are 5 crossover, 3 correction and 2 mutation options.
# In[7]:
multipoints = [1, 2, 3]
crossovers = [
{'crossover': Crossover(), 'name': 'mix'},
{'crossover': RandomCombination(), 'name': 'rnd'}]
for multipoint in multipoints:
crossover = {'crossover': UniformMultipoint(multipoint), 'name': 'uni{}'.format(multipoint)}
crossovers.append(crossover)
corrections = [
{'correction': Correction(tsp), 'name': 'vanilla'},
{'correction': MirrorCorrection(tsp), 'name': 'mirror'},
{'correction': ExtensionCorrection(tsp), 'name': 'extension'}]
parameters = [1, 3, 5]
mutations = []
for parameter in parameters:
for correction in corrections:
mutation = {'mutation': CauchyMutation(r=parameter, correction = correction['correction']), 'name': 'cauchy{}_'
.format(parameter)+correction['name']}
mutations.append(mutation)
mutation = {'mutation': LevyMutation(scale=parameter, correction = correction['correction']), 'name': 'levy{}_'
.format(parameter)+correction['name']}
mutations.append(mutation)
heur_name = 'GO_{}'.format(N)
for i in tqdm(range(num_runs), 'Testing {}'.format(heur_name)):
result = GeneticOptimization(of, maxeval, N=N, M=M, Tsel1=Tsel1, Tsel2=Tsel2,
mutation=mutation, crossover=crossover).search()
result['run'] = i
result['heur'] = heur_name
result['N'] = N
results.append(result)
return pd.DataFrame(results, columns=['heur', 'run', 'N', 'best_x', 'best_y', 'neval'])
# In[16]:
results = pd.DataFrame()
mutation = CauchyMutation(r=0.75, correction=Correction(tsp))
crossover = UniformMultipoint(1)
for N in [1, 2, 3, 5, 10, 20, 30, 100]:
res = experiment_go(of=tsp, maxeval=maxeval, num_runs=NUM_RUNS, N=N, M=N*3, Tsel1=0.1, Tsel2=1,
mutation=mutation, crossover=crossover)
results = pd.concat([results, res], axis=0)
# In[17]:
# (well-known performance criteria from previous classes)
def rel(x):
return len([n for n in x if n < np.inf])/len(x)
def mne(x):
return np.mean([n for n in x if n < np.inf])
# # Analysis
#
# **Can we improve the best configuration ($T_0=1$)?**
#
# Let's carefully analyze the data...
# In[16]:
heur = FastSimulatedAnnealing(tsp,
maxeval=1000,
T0=1,
n0=1,
alpha=2,
mutation=CauchyMutation(
r=0.5, correction=Correction(tsp)))
result = heur.search()
print('neval = {}'.format(result['neval']))
print('best_x = {}'.format(result['best_x']))
print('best_y = {}'.format(result['best_y']))
# In[17]:
log_data = result['log_data'].copy()
log_data = log_data[['step', 'x', 'f_x', 'y', 'f_y', 'T',
'swap']] # column re-ordering, for better readability
log_data.head(10)
# In[18]:
# In[12]:
results = pd.DataFrame()
for crossover in crossovers:
heur_name = 'GO_{}'.format(crossover['name'])
runs = []
for i in tqdm(range(NUM_RUNS), 'Testing {}'.format(heur_name)):
run = GeneticOptimization(tsp,
maxeval,
N=5,
M=15,
Tsel1=1,
Tsel2=0.1,
mutation=CauchyMutation(
r=.75, correction=Correction(tsp)),
crossover=crossover['crossover']).search()
run['run'] = i
run['heur'] = heur_name
run['crossover'] = crossover['name']
runs.append(run)
res_df = pd.DataFrame(
runs,
columns=['heur', 'run', 'crossover', 'best_x', 'best_y', 'neval'])
results = pd.concat([results, res_df], axis=0)
# In[13]:
def rel(x):
# In[3]:
np.random.seed(6)
# We will study DRASET's behavior on two functions - Rastrigin's function 6 and Rosenbrock's valley (De Jong's function 2). Both functions are defined here: http://www.geatbx.com/docu/fcnindex-01.html There are also graphs of these functions.
#
# Random shooting is chosen as a baseline method.
# We start with Rastrigin's function 6. This function has many local minima.
# In[4]:
function = Plato(n=2, eps=0.1)
corr = Correction(
function
) # correction is not discussed in the original article, I choose the simpliest one
# In[5]:
RUNS = 500
MAXEV = 50000 # the heuristic is implemented with maxeval parameter, I don't want to restrict heuristic by this parameter, so I set it at this high level
# We define two experiments: DRASET and random shooting.
# In[6]:
def experiment_draset(of, maxeval, num_runs, E, alfa0, N_general, N_local,
correction):
results = []
stats_fsa = stats_fsa.reset_index()
stats_fsa.sort_values(by=['alpha'])
stats_fsa
# The optimal FSA parameters are following:
# * Initial temperature : T0 = 1
# * Mutation probability: p = 0.10
# * Cooling parameters :
# - n0 = 1
# - alpha = 1
# In[28]:
mutation = BinaryMutation(p=0.10, correction=Correction(scp))
heur = FastSimulatedAnnealing(of=scp, maxeval=maxeval, T0=1, n0=1, alpha=1, mutation=mutation)
heur.reset()
result = heur.search()
print('neval = {}'.format(result['neval']))
print('best_x = {}'.format(result['best_x']))
print('best_y = {}'.format(result['best_y']))
# ## Genetic Optimization heuristic
# Let's optimize the GO heuristic in order to use the results in the mixed heuristic
# In[29]:
from heur_go import GeneticOptimization, UniformMultipoint
# optimum dj.evaluate(np.zeros(5)) # ## Generalized mutation demo on De Jong 1 # # Let's test mutation corrections first: # In[17]: from heur_aux import Correction, MirrorCorrection, ExtensionCorrection # In[18]: # sticky correction in R^n (mind x[1]) Correction(dj).correct(np.array([6.12, -4.38, 2.96])) # In[19]: # mirror correction in R^n (mind x[1]) MirrorCorrection(dj).correct(np.array([6.12, -4.38, 2.96])) # In[20]: # extension correction in R^n (mind x[1]) ExtensionCorrection(dj).correct(np.array([6.12, -4.38, 2.96])) # I.e. corrections work also in the continuous case, as expected... # In[21]:
crossovers = [
{'crossover': Crossover(), 'name': 'mix'},
{'crossover': UniformMultipoint(1), 'name': 'uni'}, # test for other n as well!
{'crossover': RandomCombination(), 'name': 'rnd'},
]
# In[12]:
results = pd.DataFrame()
for crossover in crossovers:
heur_name = 'GO_{}'.format(crossover['name'])
runs = []
for i in tqdm_notebook(range(NUM_RUNS), 'Testing {}'.format(heur_name)):
run = GeneticOptimization(tsp, maxeval, N=5, M=15, Tsel1=1.0, Tsel2=0.5,
mutation=CauchyMutation(r=1.0, correction=Correction(tsp)),
crossover=crossover['crossover']).search()
run['run'] = i
run['heur'] = heur_name
run['crossover'] = crossover['name']
runs.append(run)
res_df = pd.DataFrame(runs, columns=['heur', 'run', 'crossover', 'best_x', 'best_y', 'neval'])
results = pd.concat([results, res_df], axis=0)
# In[13]:
def rel(x):
return len([n for n in x if n < np.inf])/len(x)