def SA():
SA_iters = 10
correctCount = 0
t=0
totalTime=0
totalIters =0
global sa
sa = SimulatedAnnealing(1e11, .85, hcp)
while correctCount < NUM_RIGHT:
start = time.time()
fit = FixedIterationTrainer(sa, SA_iters)
fitness = fit.train()
t = time.time() - start
totalTime+=t
totalIters+= SA_iters
myWriter.addValue(fitness, "SA_fitness", runNum)
myWriter.addValue(t, "SA_searchTimes",runNum)
v = ef.value(sa.getOptimal())
if v == N:
correctCount += 1
else:
correctCount = 0
#SA_iters += 1
myWriter.addValue(t,"SA_times",0)
myWriter.addValue(int(SA_iters),"SA_iters",0)
print str(N) + ": SA: " + str(ef.value(sa.getOptimal())) + " took "+str(totalIters)+ " seconds and " + str(totalIters) + " iterations"
def main():
"""
Run algorithms on the gamma dataset.
Essentially ran twice for 2-fold cross validation
Metrics are evaluated outside of this file
"""
train_data = initialize_instances(TRAIN_FILE)
test_data = initialize_instances(TEST_FILE) # Get data
factory = BackPropagationNetworkFactory()
measure = SumOfSquaresError()
data_set = DataSet(train_data)
networks = [] # BackPropagationNetwork
nnop = [] # NeuralNetworkOptimizationProblem
oa = [] # OptimizationAlgorithm
oa_names = ["RHC", "SA", "GA"]
results = ""
# Create each network architecture and an optimization instance
for name in oa_names:
activation = RELU()
# Change network size
classification_network = factory.createClassificationNetwork([INPUT_LAYER, HIDDEN_LAYER1, HIDDEN_LAYER2, OUTPUT_LAYER], activation)
networks.append(classification_network)
nnop.append(NeuralNetworkOptimizationProblem(data_set, classification_network, measure))
# Randomized Optimzation Algos
oa.append(RandomizedHillClimbing(nnop[0]))
oa.append(SimulatedAnnealing(1E11, .95, nnop[1]))
oa.append(StandardGeneticAlgorithm(200, 100, 10, nnop[2]))
# Go through each optimization problem and do 2-fold CV
for i, name in enumerate(oa_names):
start = time.time()
metrics = train(oa[i], networks[i], oa_names[i], train_data, test_data, measure)
end = time.time()
training_time = end - start
results += "\nFold 1 train time: %0.03f seconds" % (training_time,)
# Write data to CSV file
with open("metrics/" + oa_names[i] + '_f1.csv', 'w') as f:
writer = csv.writer(f)
for metric in metrics:
writer.writerow(metric)
print results
# 2nd fold;
train_data = initialize_instances(TEST_FILE)
test_data = initialize_instances(TRAIN_FILE) # Get data
factory = BackPropagationNetworkFactory()
measure = SumOfSquaresError()
data_set = DataSet(train_data)
networks = [] # BackPropagationNetwork
nnop = [] # NeuralNetworkOptimizationProblem
oa = [] # OptimizationAlgorithm
oa_names = ["RHC", "SA", "GA"]
results = ""
# Create each network architecture and an optimization instance
for name in oa_names:
activation = RELU()
# Change network size
classification_network = factory.createClassificationNetwork([INPUT_LAYER, HIDDEN_LAYER1, HIDDEN_LAYER2, OUTPUT_LAYER], activation)
networks.append(classification_network)
nnop.append(NeuralNetworkOptimizationProblem(data_set, classification_network, measure))
# Randomized Optimzation Algos
oa.append(RandomizedHillClimbing(nnop[0]))
oa.append(SimulatedAnnealing(1E11, .95, nnop[1]))
oa.append(StandardGeneticAlgorithm(200, 100, 10, nnop[2]))
# Go through each optimization problem and do 2-fold CV
for i, name in enumerate(oa_names):
start = time.time()
metrics = train(oa[i], networks[i], oa_names[i], train_data, test_data, measure)
end = time.time()
training_time = end - start
results += "\nFold 1 train time: %0.03f seconds" % (training_time,)
# Write data to CSV file
with open("metrics/" + oa_names[i] + '_f2.csv', 'w') as f:
writer = csv.writer(f)
for metric in metrics:
writer.writerow(metric)
print results
network_rhc = factory.createClassificationNetwork([inputLayer, hiddenLayer, outputLayer]) nnop_rhc = NeuralNetworkOptimizationProblem(set, network_rhc, measure) rhc = RandomizedHillClimbing(nnop_rhc) fit = FixedIterationTrainer(rhc, it_rhc) fit.train() op = rhc.getOptimal(); network_rhc.setWeights(op.getData()) print "\nRHC training error:", errorRate(network_rhc, train) print "RHC training confusion matrix:", confusionMatrix(network_rhc, train) print " RHC test error:", errorRate(network_rhc, test) print " RHC test confusion matrix:", confusionMatrix(network_rhc, test) # learn weights with simulated annealing network_sa = factory.createClassificationNetwork([inputLayer, hiddenLayer, outputLayer]) nnop_sa = NeuralNetworkOptimizationProblem(set, network_sa, measure) sa = SimulatedAnnealing(1E11, 0.95, nnop_sa) fit = FixedIterationTrainer(sa, it_sa) fit.train() op = sa.getOptimal(); network_sa.setWeights(op.getData()) print "\nSA training error:", errorRate(network_sa, train) print "SA training confusion matrix:", confusionMatrix(network_sa, train) print " SA test error:", errorRate(network_sa, test) print " SA test confusion matrix:", confusionMatrix(network_sa, test) exit() # learn weights with generic algorithms network_ga = factory.createClassificationNetwork([inputLayer, hiddenLayer, outputLayer]) nnop_ga = NeuralNetworkOptimizationProblem(set, network_ga, measure) ga = StandardGeneticAlgorithm(200, 100, 10, nnop_ga)
fitness = fit.train()
results.append(ef.value(rhc.getOptimal()))
calls.append(ef.getTotalCalls())
ef.clearCount()
print "RHC, average results , " + str(sum(results)/float(runs))
print "RHC, average feval calls , " + str(sum(calls)/float(runs))
t1 = time.time() - t0
print "RHC, average time , " + str(float(t1)/runs)
t0 = time.time()
calls = []
results = []
for _ in range(runs):
sa = SimulatedAnnealing(1E11, .95, hcp)
fit = FixedIterationTrainer(sa, iters)
fitness = fit.train()
results.append(ef.value(sa.getOptimal()))
calls.append(ef.getTotalCalls())
ef.clearCount()
print "SA95, average results , " + str(sum(results)/float(runs))
print "SA95, average feval calls , " + str(sum(calls)/float(runs))
t1 = time.time() - t0
print "SA95, average time , " + str(t1/float(runs))
t0 = time.time()
calls = []
results = []
for _ in range(runs):
mf = SwapMutation()
cf = TravelingSalesmanCrossOver(ef)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, 200000)
fit.train()
print "RHC Inverse of Distance: " + str(ef.value(rhc.getOptimal()))
print "Route:"
path = []
for x in range(0, N):
path.append(rhc.getOptimal().getDiscrete(x))
print path
sa = SimulatedAnnealing(1e12, 0.999, hcp)
fit = FixedIterationTrainer(sa, 200000)
fit.train()
print "SA Inverse of Distance: " + str(ef.value(sa.getOptimal()))
print "Route:"
path = []
for x in range(0, N):
path.append(sa.getOptimal().getDiscrete(x))
print path
ga = StandardGeneticAlgorithm(2000, 1500, 250, gap)
fit = FixedIterationTrainer(ga, 1000)
fit.train()
print "GA Inverse of Distance: " + str(ef.value(ga.getOptimal()))
print "Route:"
if (max > goal):
goal = max
# run RHC
rhc = RandomizedHillClimbing(hcp)
max = 0
i = 0
while (max < goal and i < timeout):
rhc.train()
i += 1
max = ef.value(rhc.getOptimal())
#print "rhc,", i,",", max, ',', goal
print "rhc,", i,",", max, ',', goal
# run SA
sa = SimulatedAnnealing(1E11, .95, hcp)
max = 0
i = 0
while (max < goal and i < timeout):
sa.train()
i += 1
max = ef.value(sa.getOptimal())
#print "sa,", i,",", max, ',', goal
print "sa,", i,",", max, ',', goal
# run GA
ga = StandardGeneticAlgorithm(200, 100, 25, gap)
max = 0
i = 0
while (max < goal and i < timeout):
ga.train()
ef = KnapsackEvaluationFunction(weights, volumes, KNAPSACK_VOLUME, copies) odd = DiscreteUniformDistribution(ranges) nf = DiscreteChangeOneNeighbor(ranges) mf = DiscreteChangeOneMutation(ranges) cf = UniformCrossOver() df = DiscreteDependencyTree(.1, ranges) hcp = GenericHillClimbingProblem(ef, odd, nf) gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf) pop = GenericProbabilisticOptimizationProblem(ef, odd, df) rhc = RandomizedHillClimbing(hcp) fit = FixedIterationTrainer(rhc, 200000) fit.train() print "RHC: " + str(ef.value(rhc.getOptimal())) sa = SimulatedAnnealing(100, .95, hcp) fit = FixedIterationTrainer(sa, 200000) fit.train() print "SA: " + str(ef.value(sa.getOptimal())) ga = StandardGeneticAlgorithm(200, 150, 25, gap) fit = FixedIterationTrainer(ga, 1000) fit.train() print "GA: " + str(ef.value(ga.getOptimal())) mimic = MIMIC(200, 100, pop) fit = FixedIterationTrainer(mimic, 1000) fit.train() print "MIMIC: " + str(ef.value(mimic.getOptimal()))
fill = [COPIES_EACH + 1] * NUM_ITEMS
ranges = array('i', fill)
ef = KnapsackEvaluationFunction(weights, volumes, KNAPSACK_VOLUME, copies)
odd = DiscreteUniformDistribution(ranges)
nf = DiscreteChangeOneNeighbor(ranges)
mf = DiscreteChangeOneMutation(ranges)
cf = UniformCrossOver()
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
rhc = RandomizedHillClimbing(hcp)
sa = SimulatedAnnealing(SA_TEMPERATURE, SA_COOLING_FACTOR, hcp)
ga = StandardGeneticAlgorithm(GA_POPULATION, GA_CROSSOVER, GA_MUTATION,
gap)
mimic = MIMIC(MIMIC_SAMPLES, MIMIC_TO_KEEP, pop)
for n_iteration in iterations:
fit_rhc = FixedIterationTrainer(rhc, n_iteration)
fit_sa = FixedIterationTrainer(sa, n_iteration)
fit_ga = FixedIterationTrainer(ga, n_iteration)
fit_mimic = FixedIterationTrainer(mimic, n_iteration)
print("calculating the %d th iteration" % n_iteration)
# Training
start_rhc = time.time()
fit_rhc.train()
start = time.time()
fit.train()
dur = time.time() - start
print "Iters: " + str(iters) + ", Fitness: " + str(
ef.value(rhc.getOptimal())) + ", Dur: " + str(dur)
# print "Route:"
# path = []
# for x in range(0,N):
# path.append(rhc.getOptimal().getDiscrete(x))
# print path
print "Simulated Annealing"
# 1e13, 0.8, 1e12 0.85, ... dang
temp = 1E13
cooling_rate = 0.90
sa = SimulatedAnnealing(temp, cooling_rate, hcp)
for iters in iters_list:
fit = FixedIterationTrainer(sa, iters)
start = time.time()
fit.train()
dur = time.time() - start
print "Iters: " + str(iters) + ", Fitness: " + str(
ef.value(sa.getOptimal())) + ", Dur: " + str(dur)
# print "Route:"
# path = []
# for x in range(0,N):
# path.append(sa.getOptimal().getDiscrete(x))
# print path
# print "Genetic Algorithm"
# # 2000, 1500, 250 gives good results
def knapsackfunc(NUM_ITEMS, iterations):
rhcMult = 600
saMult = 600
gaMult = 4
mimicMult = 3
# Random number generator */
random = Random()
# The number of items
#NUM_ITEMS = 40
# The number of copies each
COPIES_EACH = 4
# The maximum weight for a single element
MAX_WEIGHT = 50
# The maximum volume for a single element
MAX_VOLUME = 50
# The volume of the knapsack
KNAPSACK_VOLUME = MAX_VOLUME * NUM_ITEMS * COPIES_EACH * .4
# create copies
fill = [COPIES_EACH] * NUM_ITEMS
copies = array('i', fill)
# create weights and volumes
fill = [0] * NUM_ITEMS
weights = array('d', fill)
volumes = array('d', fill)
for i in range(0, NUM_ITEMS):
weights[i] = random.nextDouble() * MAX_WEIGHT
volumes[i] = random.nextDouble() * MAX_VOLUME
# create range
fill = [COPIES_EACH + 1] * NUM_ITEMS
ranges = array('i', fill)
ef = KnapsackEvaluationFunction(weights, volumes, KNAPSACK_VOLUME, copies)
odd = DiscreteUniformDistribution(ranges)
nf = DiscreteChangeOneNeighbor(ranges)
mf = DiscreteChangeOneMutation(ranges)
cf = UniformCrossOver()
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
optimalOut = []
timeOut = []
evalsOut = []
for niter in iterations:
iterOptimalOut = [NUM_ITEMS, niter]
iterTimeOut = [NUM_ITEMS, niter]
iterEvals = [NUM_ITEMS, niter]
start = time.time()
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, niter*rhcMult)
fit.train()
end = time.time()
rhcOptimal = ef.value(rhc.getOptimal())
rhcTime = end-start
print "RHC optimum: " + str(rhcOptimal)
print "RHC time: " + str(rhcTime)
iterOptimalOut.append(rhcOptimal)
iterTimeOut.append(rhcTime)
functionEvals = ef.getNumEvals()
ef.zeroEvals()
iterEvals.append(functionEvals)
start = time.time()
sa = SimulatedAnnealing(100, .95, hcp)
fit = FixedIterationTrainer(sa, niter*saMult)
fit.train()
end = time.time()
saOptimal = ef.value(sa.getOptimal())
saTime = end-start
print "SA optimum: " + str(saOptimal)
print "SA time: " + str(saTime)
iterOptimalOut.append(saOptimal)
iterTimeOut.append(saTime)
functionEvals = ef.getNumEvals()
ef.zeroEvals()
iterEvals.append(functionEvals)
start = time.time()
ga = StandardGeneticAlgorithm(200, 150, 25, gap)
fit = FixedIterationTrainer(ga, niter*gaMult)
fit.train()
end = time.time()
gaOptimal = ef.value(ga.getOptimal())
gaTime = end - start
print "GA optimum: " + str(gaOptimal)
print "GA time: " + str(gaTime)
iterOptimalOut.append(gaOptimal)
iterTimeOut.append(gaTime)
functionEvals = ef.getNumEvals()
ef.zeroEvals()
iterEvals.append(functionEvals)
start = time.time()
mimic = MIMIC(200, 100, pop)
fit = FixedIterationTrainer(mimic, niter*mimicMult)
fit.train()
end = time.time()
mimicOptimal = ef.value(mimic.getOptimal())
mimicTime = end - start
print "MIMIC optimum: " + str(mimicOptimal)
print "MIMIC time: " + str(mimicTime)
iterOptimalOut.append(mimicOptimal)
iterTimeOut.append(mimicTime)
functionEvals = ef.getNumEvals()
ef.zeroEvals()
iterEvals.append(functionEvals)
optimalOut.append(iterOptimalOut)
timeOut.append(iterTimeOut)
evalsOut.append(iterEvals)
return [optimalOut, timeOut, evalsOut]
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
start = time.time()
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, iteration)
fit.train()
end = time.time()
rhc_time = end - start
rhc_fit = ef.value(rhc.getOptimal())
# print "RHC: " + str(rhc_fit)
start = time.time()
sa = SimulatedAnnealing(1e11, .95, hcp)
fit = FixedIterationTrainer(sa, iteration / 200)
fit.train()
end = time.time()
sa_time = end - start
sa_fit = ef.value(sa.getOptimal())
# print "SA: " + str(sa_fit)
start = time.time()
ga = StandardGeneticAlgorithm(200, 150, 25, gap)
fit = FixedIterationTrainer(ga, iteration)
fit.train()
end = time.time()
ga_time = end - start
ga_fit = ef.value(ga.getOptimal())
# print "GA: " + str(ga_fit)
def main():
"""Run algorithms on the cancer dataset."""
instances = initialize_instances()
factory = BackPropagationNetworkFactory()
measure = SumOfSquaresError()
data_set = DataSet(instances)
max_iterations = TRAINING_ITERATIONS
hidden_layer_size = HIDDEN_LAYER
# for _hidden_layer in xrange(HIDDEN_LAYER):
# hidden_layer_size = _hidden_layer + 1
network = None # BackPropagationNetwork
nnop = None # NeuralNetworkOptimizationProblem
oa = None # OptimizationAlgorithm
results = ""
for cooling in [.95, .8, .65, .5, .35, .2]:
RandomOrderFilter().filter(data_set)
train_test_split = TestTrainSplitFilter(TRAIN_TEST_SPLIT)
train_test_split.filter(data_set)
train_set = train_test_split.getTrainingSet()
test_set = train_test_split.getTestingSet()
network = factory.createClassificationNetwork(
[INPUT_LAYER, hidden_layer_size, OUTPUT_LAYER])
nnop = NeuralNetworkOptimizationProblem(train_set, network, measure)
oa = SimulatedAnnealing(1E11, cooling, nnop)
start = time.time()
correct = 0
incorrect = 0
train(oa, network, "SA", train_set, test_set, measure, cooling)
end = time.time()
training_time = end - start
optimal_instance = oa.getOptimal()
network.setWeights(optimal_instance.getData())
start = time.time()
for instance in test_set.getInstances():
network.setInputValues(instance.getData())
network.run()
predicted = instance.getLabel().getContinuous()
actual = network.getOutputValues().get(0)
if abs(predicted - actual) < 0.5:
correct += 1
else:
incorrect += 1
end = time.time()
testing_time = end - start
_results = ""
_results += "\n[SA] cooling=%0.02f" % (cooling)
_results += "\nResults for SA: \nCorrectly classified %d instances." % (
correct)
_results += "\nIncorrectly classified %d instances.\nPercent correctly classified: %0.03f%%" % (
incorrect, float(correct) / (correct + incorrect) * 100.0)
_results += "\nTraining time: %0.03f seconds" % (training_time, )
_results += "\nTesting time: %0.03f seconds\n" % (testing_time, )
with open('out/sa/cooling-%0.02f.log' % (cooling), 'w') as f:
f.write(_results)
results += _results
print results
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
# ---------------------------------------------------------------
N_ITERS = 10001
# SA
for cool_fac in [0.9, 0.92, 0.94, 0.96, 0.98]:
print cool_fac
fit_hist = []
for i in xrange(30):
hcp = GenericHillClimbingProblem(ef, odd, nf)
sa = SimulatedAnnealing(1E12, cool_fac, hcp)
fh = FixedIterTrainer(ef, sa, N_ITERS)
fit_hist.append(fh)
write_hist_csv(fit_hist, 'fitness_sa_cf_' + str(cool_fac))
for power in range(1, 12, 2):
fit_hist = []
for i in xrange(30):
hcp = GenericHillClimbingProblem(ef, odd, nf)
sa = SimulatedAnnealing(10**power, 0.94, hcp)
fh = FixedIterTrainer(ef, sa, N_ITERS)
fit_hist.append(fh)
write_hist_csv(fit_hist, 'fitness_sa_t_' + str(power))
for i in xrange(30):
hcp = GenericHillClimbingProblem(ef, odd, nf)
rhc = RandomizedHillClimbing(hcp)
fh, th = FixedIterTrainer(ef, rhc, N_ITERS)
fit_hist.append(fh)
time_hist.append(th)
write_hist_csv(fit_hist, 'fitness_rhc')
write_hist_csv(time_hist, 'time_rhc')
# SA
fit_hist = []
time_hist = []
for i in xrange(30):
hcp = GenericHillClimbingProblem(ef, odd, nf)
sa = SimulatedAnnealing(1E12, .99, hcp)
fh, th = FixedIterTrainer(ef, sa, N_ITERS)
fit_hist.append(fh)
time_hist.append(th)
write_hist_csv(fit_hist, 'fitness_sa')
write_hist_csv(time_hist, 'time_sa')
# GA
fit_hist = []
time_hist = []
for i in xrange(30):
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
ga = StandardGeneticAlgorithm(200, 100, 20, gap)
fh, th = FixedIterTrainer(ef, ga, N_ITERS)
mf = DiscreteChangeOneMutation(ranges) cf = SingleCrossOver() df = DiscreteDependencyTree(.1, ranges) hcp = GenericHillClimbingProblem(ef, odd, nf) gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf) pop = GenericProbabilisticOptimizationProblem(ef, odd, df) from time import time rhc = RandomizedHillClimbing(hcp) fit = FixedIterationTrainer(rhc, 600000) t0 = time() fit.train() print "RHC: " + str(ef.value(rhc.getOptimal())), "time taken", time() - t0 sa = SimulatedAnnealing(1E11, .95, hcp) fit = FixedIterationTrainer(sa, 600000) t0 = time() fit.train() print "SA: " + str(ef.value(sa.getOptimal())), "time taken", time() - t0 ga = StandardGeneticAlgorithm(200, 100, 10, gap) fit = FixedIterationTrainer(ga, 20000) t0 = time() fit.train() print "GA: " + str(ef.value(ga.getOptimal())), "time taken", time() - t0 mimic = MIMIC(50, 10, pop)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, N)
start = time.time()
fit.train()
end = time.time()
training_time = end - start
print "RHC Inverse of Distance: " + str(ef.value(rhc.getOptimal()))
OUTFILE = "%s%s.csv" % (OUTFILE_BASE, "RHC")
with open(OUTFILE, 'a+') as f:
f.write("%d,%f,%f\n" % (N, training_time, ef.value(rhc.getOptimal())))
sa = SimulatedAnnealing(1E12, .999, hcp)
fit = FixedIterationTrainer(sa, N)
start = time.time()
fit.train()
end = time.time()
training_time = end - start
print "SA Inverse of Distance: " + str(ef.value(sa.getOptimal()))
OUTFILE = "%s%s.csv" % (OUTFILE_BASE, "SA")
with open(OUTFILE, 'a+') as f:
f.write("%d,%f,%f\n" % (N, training_time, ef.value(sa.getOptimal())))
ga = StandardGeneticAlgorithm(2000, 1500, 250, gap)
fit = FixedIterationTrainer(ga, N)
start = time.time()
fit.train()
def main():
"""Run algorithms on the abalone dataset."""
instances = initialize_instances()
factory = BackPropagationNetworkFactory()
measure = SumOfSquaresError()
data_set = DataSet(instances)
networks = [] # BackPropagationNetwork
nnop = [] # NeuralNetworkOptimizationProblem
oa = [] # OptimizationAlgorithm
oa_names = ["RHC", "SA", "GA"]
results = ""
for name in oa_names:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER, OUTPUT_LAYER])
networks.append(classification_network)
nnop.append(
NeuralNetworkOptimizationProblem(data_set, classification_network,
measure))
oa.append(RandomizedHillClimbing(nnop[0]))
oa.append(SimulatedAnnealing(1E11, .95, nnop[1]))
oa.append(StandardGeneticAlgorithm(200, 100, 10, nnop[2]))
for i, name in enumerate(oa_names):
start = time.time()
correct = 0
incorrect = 0
train(oa[i], networks[i], oa_names[i], instances, measure)
end = time.time()
training_time = end - start
optimal_instance = oa[i].getOptimal()
networks[i].setWeights(optimal_instance.getData())
start = time.time()
for instance in instances:
networks[i].setInputValues(instance.getData())
networks[i].run()
predicted = instance.getLabel().getContinuous()
actual = networks[i].getOutputValues().get(0)
if abs(predicted - actual) < 0.5:
correct += 1
else:
incorrect += 1
end = time.time()
testing_time = end - start
results += "\nResults for %s: \nCorrectly classified %d instances." % (
name, correct)
results += "\nIncorrectly classified %d instances.\nPercent correctly classified: %0.03f%%" % (
incorrect, float(correct) / (correct + incorrect) * 100.0)
results += "\nTraining time: %0.03f seconds" % (training_time, )
results += "\nTesting time: %0.03f seconds\n" % (testing_time, )
print results
mf = SwapMutation()
cf = TravelingSalesmanCrossOver(ef)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, 200000)
fit.train()
print "RHC Inverse of Distance: " + str(ef.value(rhc.getOptimal()))
print "Route:"
path = []
for x in range(0, N):
path.append(rhc.getOptimal().getDiscrete(x))
print path
sa = SimulatedAnnealing(1E12, .999, hcp)
fit = FixedIterationTrainer(sa, 200000)
fit.train()
print "SA Inverse of Distance: " + str(ef.value(sa.getOptimal()))
print "Route:"
path = []
for x in range(0, N):
path.append(sa.getOptimal().getDiscrete(x))
print path
ga = StandardGeneticAlgorithm(2000, 1500, 250, gap)
fit = FixedIterationTrainer(ga, 1000)
fit.train()
print "GA Inverse of Distance: " + str(ef.value(ga.getOptimal()))
print "Route:"
path = []
for CE in [0.15, 0.35, 0.55, 0.75, 0.95]:
fname = outfile.replace('@ALG@', 'SA{}'.format(CE))
with open(fname, 'w') as f:
f.write('N,trial,iterations,fitness,time,fevals\n')
# Iterate over bitstring length (model complexity)
for N in range(N_min, N_max + 1, N_min):
fill = [2] * N
ranges = array('i', fill)
# Iterate over number of trials
for t in range(numTrials):
# Setup
ef = FlipFlopEvaluationFunction()
odd = DiscreteUniformDistribution(ranges)
nf = DiscreteChangeOneNeighbor(ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
sa = SimulatedAnnealing(1E10, CE, hcp)
fit = FixedIterationTrainer(sa, increments)
times = [0]
lastScore = 0
halt_count = 0
total_iter = 0
# Evaluate at iteration increments
for i in range(0, maxIters, increments):
start = clock()
fit.train()
elapsed = time.clock() - start
times.append(times[-1] + elapsed)
fevals = ef.fevals
score = ef.value(sa.getOptimal())
ef.fevals -= 1
st = '{},{},{},{},{},{},{}\n'.format(N, t, i, score, times[-1],
def main():
"""Run algorithms on the abalone dataset."""
train_instances = initialize_instances()
test_instances = initialize_instances(test=True)
factory = BackPropagationNetworkFactory()
measure = SumOfSquaresError()
data_set = DataSet(train_instances)
networks = [] # BackPropagationNetwork
oa = [] # OptimizationAlgorithm
oa_names = []
if do_rhc:
oa_names.append("RHC")
if do_sa:
oa_names.append("SA")
if do_ga:
oa_names.append("GA")
if do_bp:
oa_names.append("BP")
results = ""
# For each algo, need to see if we are doing sweeps
# No need to sweep rhc as there are no parameters
if do_rhc and sweep == False:
training_iter = TRAINING_ITERATIONS
if do_fmnist:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER, OUTPUT_LAYER])
if do_chess:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER1, HIDDEN_LAYER2, OUTPUT_LAYER])
nnop = NeuralNetworkOptimizationProblem(data_set,
classification_network,
measure)
oa = RandomizedHillClimbing(nnop)
name = "RHC"
train(oa, classification_network, name, train_instances, measure,
training_iter, test_instances, True)
if do_sa:
training_iter = TRAINING_ITERATIONS
count = 0
for temp, cooling in product(sa_temp, sa_cooling):
if do_fmnist:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER, OUTPUT_LAYER])
if do_chess:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER1, HIDDEN_LAYER2, OUTPUT_LAYER])
nnop = NeuralNetworkOptimizationProblem(data_set,
classification_network,
measure)
oa = SimulatedAnnealing(temp, cooling, nnop)
name = "SA_sweep"
if count == 0:
print_head = True
else:
print_head = False
train(oa, classification_network, name, train_instances, measure,
training_iter, test_instances, print_head, temp, cooling)
count += 1
if do_ga:
training_iter = GA_TRAINING_ITERATIONS
count = 0
for pop, prop_mate, prop_mutate in product(ga_pop, ga_prop_mate,
ga_prop_mutate):
if do_fmnist:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER, OUTPUT_LAYER])
if do_chess:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER1, HIDDEN_LAYER2, OUTPUT_LAYER])
nnop = NeuralNetworkOptimizationProblem(data_set,
classification_network,
measure)
mate = int(math.floor(pop * prop_mate))
mutate = int(math.floor(pop * prop_mutate))
oa = StandardGeneticAlgorithm(pop, mate, mutate, nnop)
name = "GA_sweep"
if count == 0:
print_head = True
else:
print_head = False
train(oa, classification_network, name, train_instances, measure,
training_iter, test_instances, print_head, pop, prop_mate,
prop_mutate)
count += 1
if do_bp and sweep == False:
training_iter = TRAINING_ITERATIONS
if do_fmnist:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER, OUTPUT_LAYER])
if do_chess:
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER1, HIDDEN_LAYER2, OUTPUT_LAYER])
oa = BatchBackPropagationTrainer(data_set, classification_network,
measure, RPROPUpdateRule())
name = "BP"
train(oa, classification_network, name, train_instances, measure,
training_iter, test_instances, True)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, N)
start = time.time()
fit.train()
end = time.time()
training_time = end - start
print "RHC: " + str(ef.value(rhc.getOptimal()))
OUTFILE = "%s%s.csv" % (OUTFILE_BASE, "RHC")
with open(OUTFILE, 'a+') as f:
f.write("%d,%f,%f\n" % (N, training_time, ef.value(rhc.getOptimal())))
sa = SimulatedAnnealing(1E11, .95, hcp)
fit = FixedIterationTrainer(sa, N)
start = time.time()
fit.train()
end = time.time()
training_time = end - start
print "SA: " + str(ef.value(sa.getOptimal()))
OUTFILE = "%s%s.csv" % (OUTFILE_BASE, "SA")
with open(OUTFILE, 'a+') as f:
f.write("%d,%f,%f\n" % (N, training_time, ef.value(sa.getOptimal())))
ga = StandardGeneticAlgorithm(200, 100, 10, gap)
fit = FixedIterationTrainer(ga, N)
start = time.time()
fit.train()
end = time.time()
def main():
N=200
tempDenom = 5
T=N/tempDenom
fill = [2] * N
ranges = array('i', fill)
iterations = 2000
gaIters = 1000
mimicIters = 1000
gaPop = 200
gaMate = 100
gaMutate = 10
mimicSamples = 200
mimicToKeep = 20
saTemp = 1E11
saCooling = .95
alg = 'all'
run = 0
settings = []
try:
opts, args = getopt.getopt(sys.argv[1:], "ahn:rsgN:m:t:i:", ["gaIters=", "mimicIters=","gaPop=", "gaMate=", "gaMutate=", "mimicSamples=", "mimicToKeep=", "saTemp=", "saCooling="])
except:
print 'knapsack.py -i <iterations> -n <NUM_ITEMS> -c <COPIES_EACH> -w <MAX_WEIGHT> -v <MAX_VOLUME>'
sys.exit(2)
for opt, arg in opts:
if opt == '-h':
print 'knapsack.py -i <iterations> -n <NUM_ITEMS> -c <COPIES_EACH> -w <MAX_WEIGHT> -v <MAX_VOLUME>'
sys.exit(1)
elif opt == '-i':
iterations = int(arg)
elif opt == '-N':
N = int(arg)
elif opt == '-t':
T = float(arg)
elif opt == '-d':
tempDenom = int(arg)
elif opt == '-r':
alg = 'RHC'
elif opt == '-a':
alg = 'all'
elif opt == '-s':
alg = 'SA'
elif opt == '-g':
alg = 'GA'
elif opt == '-m':
alg = 'MIMIC'
elif opt == '--gaPop':
gaPop = int(arg)
elif opt == '--gaMate':
gaMate = int(arg)
elif opt == '--gaMutate':
gaMutate = int(arg)
elif opt == '--mimicSamples':
mimicSamples = int(arg)
elif opt == '--mimicToKeep':
mimicToKeep = int(arg)
elif opt == '--saTemp':
saTemp = float(arg)
elif opt == '--saCooling':
saCooling = float(arg)
elif opt == '--gaIters':
gaIters = int(arg)
elif opt == '--mimicIters':
mimicIters = int(arg)
elif opt == '-n':
run = int(arg)
vars = {
'N':N,
'tempDenom':tempDenom,
'T':T,
'fill':fill,
'ranges':ranges,
'iterations' :iterations,
'gaIters':gaIters,
'mimicIters':mimicIters,
'gaPop' :gaPop,
'gaMate' :gaMate,
'gaMutate' :gaMutate,
'mimicSamples' : mimicSamples,
'mimicToKeep' : mimicToKeep,
'saTemp' : saTemp,
'saCooling' : saCooling,
'alg' : alg,
'run' : run
}
settings = getSettings(alg, settings, vars)
T=N/tempDenom
fill = [2] * N
ranges = array('i', fill)
ef = FourPeaksEvaluationFunction(T)
odd = DiscreteUniformDistribution(ranges)
nf = DiscreteChangeOneNeighbor(ranges)
mf = DiscreteChangeOneMutation(ranges)
cf = SingleCrossOver()
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
if alg == 'RHC' or alg == 'all':
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, iterations)
fit.train()
rows = []
row = []
row.append("Evaluation Function Value")
row.append(ef.value(rhc.getOptimal()))
rows.append(row)
print "RHC: " + str(ef.value(rhc.getOptimal()))
output2('4Peaks', 'RHC', rows, settings)
rows = []
buildFooter("4Peaks", "RHC", rows, settings),
outputFooter("4Peaks", "RHC", rows, settings)
if alg == 'SA' or alg == 'all':
sa = SimulatedAnnealing(saTemp, saCooling, hcp)
fit = FixedIterationTrainer(sa, iterations)
fit.train()
rows = []
row = []
row.append("Evaluation Function Value")
row.append(ef.value(sa.getOptimal()))
rows.append(row)
print "SA: " + str(ef.value(sa.getOptimal()))
output2('4Peaks', 'SA', rows, settings)
rows = []
buildFooter("4Peaks", "SA", rows, settings)
outputFooter("4Peaks", "SA", rows, settings)
if alg == 'GA' or alg == 'all':
ga = StandardGeneticAlgorithm(gaPop, gaMate, gaMutate, gap)
fit = FixedIterationTrainer(ga, gaIters)
fit.train()
print "GA: " + str(ef.value(ga.getOptimal()))
rows = []
row = []
row.append("Evaluation Function Value")
row.append(ef.value(ga.getOptimal()))
rows.append(row)
output2('4Peaks', 'GA', rows, settings)
rows = []
buildFooter("4Peaks", "GA", rows, settings)
outputFooter("4Peaks", "GA", rows , settings)
if alg == 'MIMIC' or alg == 'all':
mimic = MIMIC(mimicSamples, mimicToKeep, pop)
fit = FixedIterationTrainer(mimic, mimicIters)
fit.train()
print "MIMIC: " + str(ef.value(mimic.getOptimal()))
rows = []
row = []
row.append("Evaluation Function Value")
row.append(ef.value(mimic.getOptimal()))
rows.append(row)
output2('4Peaks', 'MIMIC', rows, settings)
rows = []
buildFooter("4Peaks", "GA", rows, settings)
outputFooter("4Peaks", "MIMIC", rows, settings)
for item in x:
stdout.write("\nRunning Four Peaks with %d iterations...\n" % item)
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, item)
start = time.time()
fit.train()
end = time.time()
value = ef.value(rhc.getOptimal())
stdout.write("RHC took %0.03f seconds and found value %d\n" % (end -
start, value))
optimal_value['RHC'].append(value)
sa = SimulatedAnnealing(1E11, .95, hcp)
fit = FixedIterationTrainer(sa, item)
start = time.time()
fit.train()
end = time.time()
value = ef.value(sa.getOptimal())
stdout.write("SA took %0.03f seconds and found value %d\n" % (end -
start, value))
optimal_value['SA'].append(value)
ga = StandardGeneticAlgorithm(200, 100, 20, gap)
fit = FixedIterationTrainer(ga, item)
start = time.time()
fit.train()
end = time.time()
value = ef.value(ga.getOptimal())
for i in range(len(iteration_list)):
iteration = iteration_list[i]
sa_total = 0
sa_time = 0
for x in range(runs):
ranges = array('i', [2] * N)
fitness = TravelingSalesmanRouteEvaluationFunction(points)
discrete_dist = DiscretePermutationDistribution(N)
discrete_neighbor = SwapNeighbor()
discrete_mutation = SwapMutation()
discrete_dependency = DiscreteDependencyTree(.1, ranges)
hill_problem = GHC(fitness, discrete_dist, discrete_neighbor)
start = time.clock()
sa_problem = SA(1E11, .95, hill_problem)
fit = FixedIterationTrainer(sa_problem, iteration)
fit.train()
end = time.clock()
full_time = end - start
sa_total += fitness.value(sa_problem.getOptimal())
sa_time += full_time
sa_total_avg = sa_total / runs
sa_time_avg = sa_time / runs
data = '{},{},{}\n'.format(iteration, sa_total_avg, sa_time_avg)
print(data)
with open(output_directory, 'a') as f:
f.write(data)
def main():
iterations = 200000
alg = 'all'
gaPop = 2000
gaMate = 1500
gaMutate = 250
mimicSamples = 500
mimicToKeep = 100
saTemp = 1E12
saCooling = .999
gaIters = 1000
mimicIters = 1000
run = 0
settings = []
try:
opts, args = getopt.getopt(sys.argv[1:], "ahrsgmn:i:", ["gaIters=", "mimicIters=", "gaPop=", "gaMate=", "gaMutate=", "mimicSamples=", "mimicToKeep=", "saTemp=", "saCooling="])
except:
print 'travelingsalesman.py -i <iterations>'
sys.exit(2)
for opt, arg in opts:
if opt == '-h':
print 'travelingsalesman.py -i <iterations>'
sys.exit(1)
elif opt == '-i':
if arg < 1:
print 'Iterations must be greater than 0'
sys.exit(2)
iterations = int(arg)
elif opt == '-a':
alg = 'all'
elif opt == '-r':
alg = 'RHC'
elif opt == '-s':
alg = 'SA'
elif opt == '-g':
alg = 'GA'
elif opt == '-m':
alg = 'MIMIC'
elif opt == '--gaPop':
if arg < 1:
print 'Population must be greater than 0'
sys.exit(2)
gaPop = int(arg)
elif opt == '--gaMate':
if arg < 1:
print 'Mating must be greater than 0'
sys.exit(2)
gaMate = int(arg)
elif opt == '--gaMutate':
if arg < 1:
print 'Mutators must be greater than 0'
sys.exit(2)
gaMutate = int(arg)
elif opt == '--mimicSamples':
if arg < 1:
print 'MIMIC samples must be greater than 0'
sys.exit(2)
mimicSamples = int(arg)
elif opt == '--mimicToKeep':
if arg < 1:
print 'MIMIC to keep must be greater than 0'
sys.exit(2)
mimicToKeep = int(arg)
elif opt == '--saTemp':
saTemp = float(arg)
elif opt == '--saCooling':
saCooling = float(arg)
elif opt == '-n':
run = int(arg)
elif opt == '--gaIters':
if arg < 1:
print 'GA Iterations must be greater than 0'
sys.exit(2)
gaIters = int(arg)
elif opt == '--mimicIters':
if arg < 1:
print 'MIMIC Iterations must be greater than 0'
sys.exit(2)
mimicIters = int(arg)
vars = {
'iterations' : iterations,
'alg' : alg,
'gaPop' : gaPop,
'gaMate' : gaMate,
'gaMutate' : gaMutate,
'mimicSamples' : mimicSamples,
'mimicToKeep' : mimicToKeep,
'saTemp' : saTemp,
'saCooling' : saCooling,
'gaIters' : gaIters,
'mimicIters' : mimicIters,
'run' : run
}
settings = getSettings(alg, settings, vars)
if gaPop < gaMate or gaPop < gaMutate or gaMate < gaMutate:
pebkac({gaPop: 'total population',gaMate : 'mating population', gaMutate : 'mutating population'}, alg, 'total population', settings)
if mimicSamples < mimicToKeep:
pebkac({mimicSamples: 'mimic samples', mimicToKeep : 'mimic to keep'}, alg, 'mimic samples', settings)
prob = 'Traveling Sales Problem'
invDist = {}
cities = CityList()
N = len(cities)
#random = Random()
points = [[0 for x in xrange(2)] for x in xrange(N)]
for i in range(0, len(points)):
coords = cities.getCoords(i)
points[i][0] = coords[0]
points[i][1] = coords[1]
ef = TravelingSalesmanRouteEvaluationFunction(points)
odd = DiscretePermutationDistribution(N)
nf = SwapNeighbor()
mf = SwapMutation()
cf = TravelingSalesmanCrossOver(ef)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
rows = []
if alg == 'RHC' or alg == 'all':
print '\n----------------------------------'
print 'Using Random Hill Climbing'
for label, setting in settings:
print label + ":" + str(setting)
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, iterations)
fit.train()
path = []
for x in range(0,N):
path.append(rhc.getOptimal().getDiscrete(x))
output(prob, 'RHC', path, points, settings)
rows = []
row = []
row.append("Inverse of Distance")
row.append(ef.value(rhc.getOptimal()))
rows.append(row)
invDist['RHC'] = ef.value(rhc.getOptimal())
buildFooter(prob, 'RHC', rows, settings)
outputFooter(prob, 'RHC', rows, settings)
if alg == 'SA' or alg == 'all':
print 'Using Simulated Annealing'
for label, setting in settings:
print label + ":" + str(setting)
sa = SimulatedAnnealing(saTemp, saCooling, hcp)
fit = FixedIterationTrainer(sa, iterations)
fit.train()
path = []
for x in range(0,N):
path.append(sa.getOptimal().getDiscrete(x))
output(prob, 'SA', path, points, settings)
rows = []
row = []
row.append("Inverse of Distance")
row.append(ef.value(sa.getOptimal()))
rows.append(row)
invDist['SA'] = ef.value(sa.getOptimal())
buildFooter(prob, 'SA', rows, settings)
outputFooter(prob, 'SA', rows, settings)
if alg == 'GA' or alg == 'all':
print '\n----------------------------------'
print 'Using Genetic Algorithm'
for label, setting in settings:
print label + ":" + str(setting)
ga = StandardGeneticAlgorithm(gaPop, gaMate, gaMutate, gap)
fit = FixedIterationTrainer(ga, gaIters)
fit.train()
path = []
for x in range(0,N):
path.append(ga.getOptimal().getDiscrete(x))
output(prob, 'GA', path, points, settings)
rows = []
row = []
row.append("Inverse of Distance")
row.append(ef.value(ga.getOptimal()))
rows.append(row)
invDist['GA'] = ef.value(ga.getOptimal())
buildFooter(prob, 'GA', rows, settings)
outputFooter(prob, 'GA', rows, settings)
if alg == 'MIMIC' or alg == 'all':
print '\n----------------------------------'
print 'Using MIMIC'
for label, setting in settings:
print label + ":" + str(setting)
# for mimic we use a sort encoding
ef = TravelingSalesmanSortEvaluationFunction(points);
fill = [N] * N
ranges = array('i', fill)
odd = DiscreteUniformDistribution(ranges);
df = DiscreteDependencyTree(.1, ranges);
pop = GenericProbabilisticOptimizationProblem(ef, odd, df);
mimic = MIMIC(mimicSamples, mimicToKeep, pop)
fit = FixedIterationTrainer(mimic, mimicIters)
fit.train()
path = []
optimal = mimic.getOptimal()
fill = [0] * optimal.size()
ddata = array('d', fill)
for i in range(0,len(ddata)):
ddata[i] = optimal.getContinuous(i)
order = ABAGAILArrays.indices(optimal.size())
ABAGAILArrays.quicksort(ddata, order)
output(prob, 'MIMIC', order, points, settings)
rows = []
row = []
row.append("Inverse of Distance")
row.append(ef.value(mimic.getOptimal()))
rows.append(row)
invDist['MIMIC'] = ef.value(mimic.getOptimal())
buildFooter(prob, 'MIMIC', rows, settings)
outputFooter(prob, 'MIMIC', rows, settings)
maxn = max(len(key) for key in invDist)
maxd = max(len(str(invDist[key])) for key in invDist)
print "Results"
for result in invDist:
print "%-*s %s %-*s" % (len('Best Alg') + 2, result, ':', maxd, invDist[result])
if alg == 'all':
print "%-*s %s %-*s" % (len('Best Alg') + 2, 'Best Alg', ':', maxd, max(invDist.iterkeys(), key=(lambda key: invDist[key])))
print '----------------------------------'
def main():
trainingInstances, testingInstances = initialize_instances()
factory = BackPropagationNetworkFactory()
measure = SumOfSquaresError()
data_set = DataSet(trainingInstances)
networks = [] # BackPropagationNetwork
nnop = [] # NeuralNetworkOptimizationProblem
oa = [] # OptimizationAlgorithm
oa_names = ["RHC", "SA_15", "SA_35", "SA_55", "SA_75", "SA_95"]
#oa_names=["GA_100_50_5", "GA_200_50_5", "GA_100_50_10", "GA_200_50_10", "GA_100_100_5", "GA_200_100_5", "GA_100_100_10", "GA_200_100_10"]
#oa_names=["GA_200_100_5", "GA_100_100_10", "GA_200_100_10"]
for name in oa_names:
#use RELU activation function
classification_network = factory.createClassificationNetwork(
[INPUT_LAYER, HIDDEN_LAYER, OUTPUT_LAYER], ReLU())
networks.append(classification_network)
nnop.append(
NeuralNetworkOptimizationProblem(data_set, classification_network,
measure))
oa.append(RandomizedHillClimbing(nnop[0]))
oa.append(SimulatedAnnealing(1E11, .15, nnop[1]))
oa.append(SimulatedAnnealing(1E11, .35, nnop[2]))
oa.append(SimulatedAnnealing(1E11, .55, nnop[3]))
oa.append(SimulatedAnnealing(1E11, .75, nnop[4]))
oa.append(SimulatedAnnealing(1E11, .95, nnop[5]))
# oa.append(StandardGeneticAlgorithm(100, 50, 5, nnop[0]))
# oa.append(StandardGeneticAlgorithm(200, 50, 5, nnop[1]))
# oa.append(StandardGeneticAlgorithm(100, 50, 10, nnop[2]))
# oa.append(StandardGeneticAlgorithm(200, 50, 10, nnop[3]))
# oa.append(StandardGeneticAlgorithm(100, 100, 5, nnop[4]))
#oa.append(StandardGeneticAlgorithm(200, 100, 5, nnop[0]))
#oa.append(StandardGeneticAlgorithm(100, 100, 10, nnop[1]))
#oa.append(StandardGeneticAlgorithm(200, 100, 10, nnop[2]))
with open('nn_spam_results_RHC_SA.csv', 'w') as csvfile:
writer = csv.writer(csvfile)
for i, name in enumerate(oa_names):
results = ''
start = time.time()
traincorrect = 0
trainincorrect = 0
testcorrect = 0
testincorrect = 0
train(oa[i], networks[i], oa_names[i], trainingInstances,
testingInstances, measure)
end = time.time()
training_time = end - start
optimal_instance = oa[i].getOptimal()
networks[i].setWeights(optimal_instance.getData())
start = time.time()
for instance in trainingInstances:
networks[i].setInputValues(instance.getData())
networks[i].run()
predicted = instance.getLabel().getContinuous()
actual = networks[i].getOutputValues().get(0)
if abs(predicted - actual) < 0.5:
traincorrect += 1
else:
trainincorrect += 1
for instance in testingInstances:
networks[i].setInputValues(instance.getData())
networks[i].run()
predicted = instance.getLabel().getContinuous()
actual = networks[i].getOutputValues().get(0)
if abs(predicted - actual) < 0.5:
testcorrect += 1
else:
testincorrect += 1
end = time.time()
testing_time = end - start
results += "\nResults for %s: \nCorrectly classified %d training instances." % (
name, traincorrect)
results += "\nIncorrectly classified %d instances.\nPercent correctly classified: %0.03f%%" % (
trainincorrect, float(traincorrect) /
(traincorrect + trainincorrect) * 100.0)
results += "\nResults for %s: \nCorrectly classified %d testing instances." % (
name, testcorrect)
results += "\nIncorrectly classified %d instances.\nPercent correctly classified: %0.03f%%" % (
testincorrect, float(testcorrect) /
(testcorrect + testincorrect) * 100.0)
results += "\nTraining time: %0.03f seconds" % (training_time, )
results += "\nTesting time: %0.03f seconds\n" % (testing_time, )
print results
writer.writerow([results])
writer.writerow('')
ef = ContinuousPeaksEvaluationFunction(T) odd = DiscreteUniformDistribution(ranges) nf = DiscreteChangeOneNeighbor(ranges) mf = DiscreteChangeOneMutation(ranges) cf = SingleCrossOver() df = DiscreteDependencyTree(.1, ranges) hcp = GenericHillClimbingProblem(ef, odd, nf) gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf) pop = GenericProbabilisticOptimizationProblem(ef, odd, df) rhc = RandomizedHillClimbing(hcp) fit = FixedIterationTrainer(rhc, 200000) fit.train() print "RHC: " + str(ef.value(rhc.getOptimal())) sa = SimulatedAnnealing(1E11, .95, hcp) fit = FixedIterationTrainer(sa, 200000) fit.train() print "SA: " + str(ef.value(sa.getOptimal())) ga = StandardGeneticAlgorithm(200, 100, 10, gap) fit = FixedIterationTrainer(ga, 1000) fit.train() print "GA: " + str(ef.value(ga.getOptimal())) mimic = MIMIC(200, 20, pop) fit = FixedIterationTrainer(mimic, 1000) fit.train() print "MIMIC: " + str(ef.value(mimic.getOptimal()))
calls = []
results = []
for _ in range(runs):
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, iters)
fitness = fit.train()
results.append(ef.value(rhc.getOptimal()))
calls.append(ef.getTotalCalls())
ef.clearCount()
print "RHC, average results , " + str(sum(results)/float(runs)) + ", fourpeaks%d.txt" % N
print "RHC, average feval calls , " + str(sum(calls)/float(runs)) + ", fourpeaks%d.txt" % N
t1 = time.time() - t0
print "RHC, average time , " + str(float(t1)/runs) + ", fourpeaks%d.txt" % N
t0 = time.time()
calls = []
results = []
for _ in range(runs):
sa = SimulatedAnnealing(1e10, .95, hcp)
fit = FixedIterationTrainer(sa, iters)
fitness = fit.train()
results.append(ef.value(sa.getOptimal()))
calls.append(ef.getTotalCalls())
ef.clearCount()
print "SA, average results , " + str(sum(results)/float(runs)) + ", fourpeaks%d.txt" % N
print "SA, average feval calls , " + str(sum(calls)/float(runs)) + ", fourpeaks%d.txt" % N
t1 = time.time() - t0
print "SA, average time , " + str(t1/float(runs)) + ", fourpeaks%d.txt" % N
def main():
# The number of items
NUM_ITEMS = 40
# The number of copies each
COPIES_EACH = 4
# The maximum weight for a single element
MAX_WEIGHT = 50
# The maximum volume for a single element
MAX_VOLUME = 50
iterations = 20000
gaIters = 1000
mimicIters = 1000
gaPop = 200
gaMate = 150
gaMutate = 25
mimicSamples = 200
mimicToKeep = 100
saTemp = 100
saCooling = .95
alg = 'all'
run = 0
settings = []
try:
opts, args = getopt.getopt(sys.argv[1:], "ahrsgmn:N:c:w:v:i:", ["gaIters=", "mimicIters=","gaPop=", "gaMate=", "gaMutate=", "mimicSamples=", "mimicToKeep=", "saTemp=", "saCooling="])
except:
print 'knapsack.py -i <iterations> -n <NUM_ITEMS> -c <COPIES_EACH> -w <MAX_WEIGHT> -v <MAX_VOLUME>'
sys.exit(2)
for opt, arg in opts:
if opt == '-h':
print 'knapsack.py -i <iterations> -n <NUM_ITEMS> -c <COPIES_EACH> -w <MAX_WEIGHT> -v <MAX_VOLUME>'
sys.exit(1)
elif opt == '-i':
iterations = int(arg)
elif opt == '-N':
NUM_ITEMS = int(arg)
elif opt == '-c':
COPIES_EACH = int(arg)
elif opt == '-w':
MAX_WEIGHT = int(arg)
elif opt == '-v':
MAX_VOLUME = int(arg)
elif opt == '-n':
run = int(arg)
elif opt == '-r':
alg = 'RHC'
elif opt == '-s':
alg = 'SA'
elif opt == '-g':
alg = 'GA'
elif opt == '-m':
alg = 'MIMIC'
elif opt == '-a':
alg = 'all'
elif opt == '--gaPop':
gaPop = int(arg)
elif opt == '--gaMate':
gaMate = int(arg)
elif opt == '--gaMutate':
gaMutate = int(arg)
elif opt == '--mimicSamples':
mimicSamples = int(arg)
elif opt == '--mimicToKeep':
mimicToKeep = int(arg)
elif opt == '--saTemp':
saTemp = float(arg)
elif opt == '--saCooling':
saCooling = float(arg)
elif opt == '--gaIters':
gaIters = int(arg)
elif opt == '--mimicIters':
mimicIters = int(arg)
vars ={
'NUM_ITEMS' : NUM_ITEMS,
'COPIES_EACH' : COPIES_EACH,
'MAX_WEIGHT' : MAX_WEIGHT,
'MAX_VOLUME' : MAX_VOLUME,
'iterations' : iterations,
'gaIters' : gaIters,
'mimicIters' : mimicIters,
'gaPop' : gaPop,
'gaMate' : gaMate,
'gaMutate' : gaMutate,
'mimicSamples' : mimicSamples,
'mimicToKeep' : mimicToKeep,
'saTemp' : saTemp,
'saCooling' : saCooling,
'alg' : alg,
'run' : run
}
settings = getSettings(alg, settings, vars)
# Random number generator */
random = Random()
# The volume of the knapsack
KNAPSACK_VOLUME = MAX_VOLUME * NUM_ITEMS * COPIES_EACH * .4
# create copies
fill = [COPIES_EACH] * NUM_ITEMS
copies = array('i', fill)
# create weights and volumes
fill = [0] * NUM_ITEMS
weights = array('d', fill)
volumes = array('d', fill)
for i in range(0, NUM_ITEMS):
weights[i] = random.nextDouble() * MAX_WEIGHT
volumes[i] = random.nextDouble() * MAX_VOLUME
# create range
fill = [COPIES_EACH + 1] * NUM_ITEMS
ranges = array('i', fill)
ef = KnapsackEvaluationFunction(weights, volumes, KNAPSACK_VOLUME, copies)
odd = DiscreteUniformDistribution(ranges)
nf = DiscreteChangeOneNeighbor(ranges)
mf = DiscreteChangeOneMutation(ranges)
cf = UniformCrossOver()
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
if alg == 'RHC' or alg == 'all':
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, iterations)
fit.train()
print "RHC: " + str(ef.value(rhc.getOptimal()))
rows = []
row = []
row.append("Evaluation Function Value")
row.append(str(ef.value(rhc.getOptimal())))
rows.append(row)
output2('Knapsack', 'RHC', rows, settings)
rows = []
buildFooter("Knapsack", "RHC", rows, settings)
outputFooter("Knapsack", "RHC", rows , settings)
if alg == 'SA' or alg == 'all':
sa = SimulatedAnnealing(saTemp, saCooling, hcp)
fit = FixedIterationTrainer(sa, iterations)
fit.train()
rows = []
row = []
row.append("Evaluation Function Value")
row.append(ef.value(sa.getOptimal()))
rows.append(row)
print "SA: " + str(ef.value(sa.getOptimal()))
output2('Knapsack', 'SA', rows, settings)
rows = []
buildFooter("Knapsack", "SA", rows, settings)
outputFooter("Knapsack", "SA", rows, settings)
if alg == 'GA' or alg == 'all':
ga = StandardGeneticAlgorithm(gaPop, gaMate, gaMutate, gap)
fit = FixedIterationTrainer(ga, gaIters)
fit.train()
rows = []
row = []
row.append("Evaluation Function Value")
row.append(ef.value(ga.getOptimal()))
rows.append(row)
print "GA: " + str(ef.value(ga.getOptimal()))
output2('Knapsack', 'GA', rows, settings)
buildFooter("Knapsack", "GA", rows, settings)
outputFooter("Knapsack", "GA", rows , settings)
if alg == 'MIMIC' or alg == 'all':
mimic = MIMIC(mimicSamples, mimicToKeep, pop)
fit = FixedIterationTrainer(mimic, mimicIters)
fit.train()
print "MIMIC: " + str(ef.value(mimic.getOptimal()))
rows = []
row = []
row.append("Evaluation Function Value")
row.append(ef.value(mimic.getOptimal()))
rows.append(row)
output2('Knapsack', 'MIMIC', rows, settings)
rows = []
buildFooter("Knapsack", "MIMIC", rows, settings)
outputFooter("Knapsack", "MIMIC", rows , settings)
def run_all_2(N=200, T=40, fout=None):
problem = 'fourpeaks'
# N=200
# T=N/10
maxEpochs = 10**6
maxTime = 300 #5 minutes
fill = [2] * N
ranges = array('i', fill)
ef = FourPeaksEvaluationFunction(T)
odd = DiscreteUniformDistribution(ranges)
nf = DiscreteChangeOneNeighbor(ranges)
mf = DiscreteChangeOneMutation(ranges)
# mf = SwapMutation()
cf = SingleCrossOver()
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)
def run_algo(alg, fit, label, difficulty, iters):
trainTimes = [0.]
trainTime = []
scoreChange = [0.]
stuckCount = 10**3
prev = 0.
for epoch in range(0, maxEpochs, 1):
st = time.clock()
fit.train()
et = time.clock()
trainTimes.append(trainTimes[-1] + (et - st))
trainTime.append((et - st))
rollingMean = 10
avgTime = (math.fsum(trainTime[-rollingMean:]) /
float(rollingMean))
score = ef.value(alg.getOptimal())
# trialString = '{}-{}-{}-{}'.format(label,score,epoch,trainTimes[-1])
trialData = [
problem, difficulty, label, score, epoch, trainTimes[-1],
avgTime, iters
]
# print(trialData)
# fout.writerow(trialData)
# print(trialData)
print(trialData, max(scoreChange))
# print(max(scoreChange))
optimum = (difficulty - 1 - T) + difficulty
if score >= optimum: break
scoreChange.append(abs(score - prev))
prev = score
scoreChange = scoreChange[-stuckCount:]
# print(scoreChange)
if max(scoreChange) == 0: break
if trainTimes[-1] > maxTime: break
# print(trialData)
fout.writerow(trialData)
iters = 1000
rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, iters)
run_algo(rhc, fit, 'RHC', N, iters)
iters = 1000
startTemp = 1E10
coolingFactor = .95
sa = SimulatedAnnealing(startTemp, coolingFactor, hcp)
fit = FixedIterationTrainer(sa, iters)
run_algo(sa, fit, 'SA', N, iters)
iters = 10
population = 300
mates = 100
mutations = 50
ga = StandardGeneticAlgorithm(population, mates, mutations, gap)
fit = FixedIterationTrainer(ga, iters)
run_algo(ga, fit, 'GA', N, iters)
iters = 10
samples = 200
keep = 20
mimic = MIMIC(samples, keep, pop)
fit = FixedIterationTrainer(mimic, iters)
run_algo(mimic, fit, 'MIMIC', N, iters)