class XCSProgram:
def __init__(self):
self.env = XCSEnvironment()
self.perf = []
def init(self):
self.env = XCSEnvironment()
self.perf = []
def run_experiments(self):
for exp in range(conf.max_experiments):
random.seed(exp)
self.actual_time = 0.0
self.pop = XCSClassifierSet(self.env,self.actual_time)
self.init()
for iteration in range(conf.max_iterations):
self.run_explor()
self.run_exploit(iteration)
print "now" + str(exp)
self.file_writer(exp)
self.performance_writer(exp)
self.make_graph()
def run_explor(self):
"""環境の状態をセット"""
self.env.set_state()
"""MatchSet[M]を生成"""
self.match_set = XCSMatchSet(self.pop,self.env,self.actual_time)
"""MatchSet[M]に基いて,prediction array[PA]を生成"""
self.generate_prediction_array()
"""prediction array[PA]に基づいて行動選択"""
self.select_action()
"""選択した行動に基いて,ActionSet[A]を生成する."""
self.action_set = XCSActionSet(self.match_set,self.action,self.env,self.actual_time)
"""行動を取る. 強化学習が働く"""
self.action_set.do_action()
"""ActionSet[A]内に対してパラメータ更新"""
self.action_set.update_action_set()
"""ActionSet[A]内でルールの包摂をする"""
self.action_set.do_action_set_subsumption(self.pop)
"""ActionSet[A]に対してGAを回す."""
self.run_GA()
if len(self.pop.cls) > conf.N:
self.pop.delete_from_population()
self.actual_time += 1.0
def run_exploit(self,iteration):
if iteration%100==0:
p = 0
for i in range(100):
self.env.set_state()
self.match_set = XCSMatchSet(self.pop,self.env,self.actual_time)
self.generate_prediction_array()
self.action = self.best_action()
if self.env.is_true(self.action):
p += 1
self.perf.append(p)
def select_action(self):
if random.random() > conf.p_explr:
self.action = self.best_action()
else:
self.action = random.randrange(2)
def best_action(self):
big = self.p_array[0]
best = 0
for i in range(2):
if big < self.p_array[i]:
big = self.p_array[i]
best = i
return best
def generate_prediction_array(self):
self.p_array = [0,0]
self.f_array = [0,0]
for cl in self.match_set.cls:
self.p_array[cl.action] += cl.prediction*cl.fitness
self.f_array[cl.action] += cl.fitness
for i in range(2):
if self.f_array[i] != 0:
self.p_array[i] /= self.f_array[i]
def select_offspring(self):
"""fitnessを元に親をルーレット選択"""
fit_sum = self.action_set.fitness_sum()
choice_point = fit_sum * random.random()
fit_sum = 0.0
for cl in self.action_set.cls:
fit_sum += cl.fitness
if fit_sum > choice_point:
return cl
return None
def apply_crossover(self,cl1,cl2):
"""2点交叉適用"""
length = len(cl1.condition)
sep1 = int(random.random()*(length))
sep2 = int(random.random()*(length))
if sep1>sep2:
sep1,sep2 = sep2,sep1
elif sep1==sep2:
sep2 = sep2+1
cond1 = cl1.condition
cond2 = cl2.condition
for i in range(sep1,sep2):
if cond1[i] != cond2[i]:
cond1[i],cond2[i] = cond2[i],cond1[i]
cl1.condition = cond1
cl2.condition = cond2
def apply_mutation(self,cl):
"""突然変異"""
i = 0
for i in range(len(cl.condition)):
if random.random() < conf.myu:
if cl.condition[i] == '#':
cl.condition[i] = self.env.state[i]
else:
cl.condition[i] = '#'
if random.random() < conf.myu:
cl.action = random.randrange(2)
def run_GA(self):
if self.actual_time - self.action_set.ts_num_sum()/self.action_set.numerosity_sum()>conf.theta_ga:
for cl in self.action_set.cls:
cl.time_stamp = self.actual_time
parent1 = self.select_offspring()
parent2 = self.select_offspring()
child1 = parent1.deep_copy(self.actual_time)
child2 = parent2.deep_copy(self.actual_time)
child1.numerosity = 1
child2.numerosity = 1
child1.experience = 0
child2.experience = 0
if random.random() < conf.chi:
self.apply_crossover(child1,child2)
child1.prediction = (parent1.prediction+parent2.prediction)/2.0
child1.error = 0.25*(parent1.error+parent2.error)/2.0
child1.fitness = 0.1*(parent1.fitness+parent2.fitness)/2.0
child2.prediction = child1.prediction
child2.error = child1.error
child2.fitness = child1.fitness
self.apply_mutation(child1)
self.apply_mutation(child2)
if conf.doGASubsumption:
if parent1.does_subsume(child1):
parent1.numerosity += 1
elif parent2.does_subsume(child1):
parent2.numerosity += 1
else:
self.pop.insert_in_population(child1)
if parent1.does_subsume(child2):
parent1.numerosity += 1
elif parent2.does_subsume(child2):
parent2.numerosity += 1
else:
self.pop.insert_in_population(child2)
else:
self.pop.insert_in_population(child1)
self.pop.insert_in_population(child2)
while self.pop.numerosity_sum() > conf.N:
self.pop.delete_from_population()
def file_writer(self,num):
file_name = "population"+str(num)+".csv"
write_csv = csv.writer(file(file_name,'w'),lineterminator='\n')
write_csv.writerow(["condition","action","fitness","prediction","error","numerosity","experience","time_stamp","action_set_size"])
for cl in self.pop.cls:
cond = ""
for c in cl.condition:
cond += str(c)
write_csv.writerow([cond,cl.action,cl.fitness,cl.prediction,cl.error,cl.numerosity,cl.experience,cl.time_stamp,cl.action_set_size])
def performance_writer(self,num):
file_name = "performance" + str(num) + ".csv"
np.savetxt(file_name, np.array(self.perf),fmt="%d", delimiter=",")
def make_graph(self):
performance = []
"""操作するファイルはperformance0.csvスタート"""
i = 0
file_path = "performance" + str(i) + ".csv"
while os.path.exists(file_path):
pf = np.loadtxt(file_path,delimiter=",")
performance.append(pf)
i += 1
file_path = "performance" + str(i) + ".csv"
"""データの数 = whileでインクリメントした分"""
data_num = i
"""データの中身の長さ = np.loadtxtしたデータのlen"""
data_length = len(np.loadtxt("performance0.csv",delimiter=","))
pf = []
"""0, 100, 200, 300, ..., data_length*100"""
x = np.arange(0,data_length*100,100)
for i in range(data_length):
sum = 0.0
for j in range(data_num):
sum += performance[j][i]
pf.append(sum/float(data_num))
pf = np.array(pf)
np.savetxt("ave_performance.csv",pf,delimiter=",")
fig = plt.figure(figsize=(16, 10))
ax = fig.add_subplot(1,1,1)
ax.plot(x, pf, linewidth=2, label='performance')
ax.set_ylim(40, 110)
ax.set_xlim(0, data_length*100)
ax.set_title('Performance')
ax.set_yticklabels(['40%','50%','60%','70%','80%','90%','100%',''])
ax.grid()
filenamepng = "performance.png"
plt.savefig(filenamepng, dpi=150)
filenameeps = "performance.eps"
plt.savefig(filenameeps)
plt.show()
class XCSProgram:
def __init__(self):
self.env = XCSEnvironment()
self.perf = []
def init(self):
self.env = XCSEnvironment()
self.perf = []
def run_experiments(self):
for exp in range(conf.max_experiments):
random.seed(exp)
self.actual_time = 0.0
self.pop = XCSClassifierSet(self.env, self.actual_time)
self.init()
for iteration in range(conf.max_iterations):
self.run_explor()
self.run_exploit(iteration)
print "now" + str(exp)
self.file_writer(exp)
self.performance_writer(exp)
self.make_graph()
def run_explor(self):
"""環境の状態をセット"""
self.env.set_state()
"""MatchSet[M]を生成"""
self.match_set = XCSMatchSet(self.pop, self.env, self.actual_time)
"""MatchSet[M]に基いて,prediction array[PA]を生成"""
self.generate_prediction_array()
"""prediction array[PA]に基づいて行動選択"""
self.select_action()
"""選択した行動に基いて,ActionSet[A]を生成する."""
self.action_set = XCSActionSet(self.match_set, self.action, self.env,
self.actual_time)
"""行動を取る. 強化学習が働く"""
self.action_set.do_action()
"""ActionSet[A]内に対してパラメータ更新"""
self.action_set.update_action_set()
"""ActionSet[A]内でルールの包摂をする"""
self.action_set.do_action_set_subsumption(self.pop)
"""ActionSet[A]に対してGAを回す."""
self.run_GA()
if len(self.pop.cls) > conf.N:
self.pop.delete_from_population()
self.actual_time += 1.0
def run_exploit(self, iteration):
if iteration % 100 == 0:
p = 0
for i in range(100):
self.env.set_state()
self.match_set = XCSMatchSet(self.pop, self.env,
self.actual_time)
self.generate_prediction_array()
self.action = self.best_action()
if self.env.is_true(self.action):
p += 1
self.perf.append(p)
def select_action(self):
if random.random() > conf.p_explr:
self.action = self.best_action()
else:
self.action = random.randrange(2)
def best_action(self):
big = self.p_array[0]
best = 0
for i in range(2):
if big < self.p_array[i]:
big = self.p_array[i]
best = i
return best
def generate_prediction_array(self):
self.p_array = [0, 0]
self.f_array = [0, 0]
for cl in self.match_set.cls:
self.p_array[cl.action] += cl.prediction * cl.fitness
self.f_array[cl.action] += cl.fitness
for i in range(2):
if self.f_array[i] != 0:
self.p_array[i] /= self.f_array[i]
def select_offspring(self):
"""fitnessを元に親をルーレット選択"""
fit_sum = self.action_set.fitness_sum()
choice_point = fit_sum * random.random()
fit_sum = 0.0
for cl in self.action_set.cls:
fit_sum += cl.fitness
if fit_sum > choice_point:
return cl
return None
def apply_crossover(self, cl1, cl2):
"""2点交叉適用"""
length = len(cl1.condition)
sep1 = int(random.random() * (length))
sep2 = int(random.random() * (length))
if sep1 > sep2:
sep1, sep2 = sep2, sep1
elif sep1 == sep2:
sep2 = sep2 + 1
cond1 = cl1.condition
cond2 = cl2.condition
for i in range(sep1, sep2):
if cond1[i] != cond2[i]:
cond1[i], cond2[i] = cond2[i], cond1[i]
cl1.condition = cond1
cl2.condition = cond2
def apply_mutation(self, cl):
"""突然変異"""
i = 0
for i in range(len(cl.condition)):
if random.random() < conf.myu:
if cl.condition[i] == '#':
cl.condition[i] = self.env.state[i]
else:
cl.condition[i] = '#'
if random.random() < conf.myu:
cl.action = random.randrange(2)
def run_GA(self):
if self.actual_time - self.action_set.ts_num_sum(
) / self.action_set.numerosity_sum() > conf.theta_ga:
for cl in self.action_set.cls:
cl.time_stamp = self.actual_time
parent1 = self.select_offspring()
parent2 = self.select_offspring()
child1 = parent1.deep_copy(self.actual_time)
child2 = parent2.deep_copy(self.actual_time)
child1.numerosity = 1
child2.numerosity = 1
child1.experience = 0
child2.experience = 0
if random.random() < conf.chi:
self.apply_crossover(child1, child2)
child1.prediction = (parent1.prediction +
parent2.prediction) / 2.0
child1.error = 0.25 * (parent1.error + parent2.error) / 2.0
child1.fitness = 0.1 * (parent1.fitness +
parent2.fitness) / 2.0
child2.prediction = child1.prediction
child2.error = child1.error
child2.fitness = child1.fitness
self.apply_mutation(child1)
self.apply_mutation(child2)
if conf.doGASubsumption:
if parent1.does_subsume(child1):
parent1.numerosity += 1
elif parent2.does_subsume(child1):
parent2.numerosity += 1
else:
self.pop.insert_in_population(child1)
if parent1.does_subsume(child2):
parent1.numerosity += 1
elif parent2.does_subsume(child2):
parent2.numerosity += 1
else:
self.pop.insert_in_population(child2)
else:
self.pop.insert_in_population(child1)
self.pop.insert_in_population(child2)
while self.pop.numerosity_sum() > conf.N:
self.pop.delete_from_population()
def file_writer(self, num):
file_name = "population" + str(num) + ".csv"
write_csv = csv.writer(file(file_name, 'w'), lineterminator='\n')
write_csv.writerow([
"condition", "action", "fitness", "prediction", "error",
"numerosity", "experience", "time_stamp", "action_set_size"
])
for cl in self.pop.cls:
cond = ""
for c in cl.condition:
cond += str(c)
write_csv.writerow([
cond, cl.action, cl.fitness, cl.prediction, cl.error,
cl.numerosity, cl.experience, cl.time_stamp, cl.action_set_size
])
def performance_writer(self, num):
file_name = "performance" + str(num) + ".csv"
np.savetxt(file_name, np.array(self.perf), fmt="%d", delimiter=",")
def make_graph(self):
performance = []
"""操作するファイルはperformance0.csvスタート"""
i = 0
file_path = "performance" + str(i) + ".csv"
while os.path.exists(file_path):
pf = np.loadtxt(file_path, delimiter=",")
performance.append(pf)
i += 1
file_path = "performance" + str(i) + ".csv"
"""データの数 = whileでインクリメントした分"""
data_num = i
"""データの中身の長さ = np.loadtxtしたデータのlen"""
data_length = len(np.loadtxt("performance0.csv", delimiter=","))
pf = []
"""0, 100, 200, 300, ..., data_length*100"""
x = np.arange(0, data_length * 100, 100)
for i in range(data_length):
sum = 0.0
for j in range(data_num):
sum += performance[j][i]
pf.append(sum / float(data_num))
pf = np.array(pf)
np.savetxt("ave_performance.csv", pf, delimiter=",")
fig = plt.figure(figsize=(16, 10))
ax = fig.add_subplot(1, 1, 1)
ax.plot(x, pf, linewidth=2, label='performance')
ax.set_ylim(40, 110)
ax.set_xlim(0, data_length * 100)
ax.set_title('Performance')
ax.set_yticklabels(
['40%', '50%', '60%', '70%', '80%', '90%', '100%', ''])
ax.grid()
filenamepng = "performance.png"
plt.savefig(filenamepng, dpi=150)
filenameeps = "performance.eps"
plt.savefig(filenameeps)
plt.show()
