def estimate_log_evidence_pm_pendulum(load_path="./result_apply/",
save_path="./result_apply/",
loop_num=0,
i=5):
gather_sample.gather_real_world_sample(loop_num=loop_num)
X = np.loadtxt(load_path + 'real_world_samples_input.csv', delimiter=',')
Y = np.loadtxt(load_path + 'real_world_samples_output.csv', delimiter=',')
X = np.delete(X, i, 0)
Y = np.delete(Y, i, 0)
pm_class = pendulum_pm.PendulumPM(X, Y)
gather_sample.gather_real_world_sample(loop_num=loop_num)
X = np.loadtxt(load_path + 'real_world_samples_input.csv', delimiter=',')
Y = np.loadtxt(load_path + 'real_world_samples_output.csv', delimiter=',')
X = np.delete(X, i, 0)
Y = np.delete(Y, i, 0)
X2 = np.loadtxt(load_path + 'real_world_samples_input.csv', delimiter=',')
Y2 = np.loadtxt(load_path + 'real_world_samples_output.csv', delimiter=',')
print("prev_log_evidence =", pm_class.log_evidence())
print("debug_log_evidence =", pm_class.log_evidence_of_new_data(X, Y))
print("new_log_evidence =", pm_class.log_evidence_of_new_data(X2, Y2))
print("diff =",
pm_class.log_evidence_of_new_data(X2, Y2) - pm_class.log_evidence())
temp_data = np.array(
[pm_class.log_evidence(),
pm_class.log_evidence_of_new_data(X2, Y2)])
return (pm_class.log_evidence_of_new_data(X2, Y2) -
pm_class.log_evidence())
def estimate_log_evidence_pm_pendulum(load_path="./result_apply/",
save_path="./result_apply/",
loop_num=0):
gather_sample.gather_real_world_sample(loop_num=loop_num)
X = np.loadtxt(load_path+'real_world_samples_input.csv', delimiter=',')
Y = np.loadtxt(load_path+'real_world_samples_output.csv', delimiter=',')
pm_class = pendulum_pm.PendulumPM(X,Y)
gather_sample.gather_real_world_sample(loop_num=loop_num)
X = np.loadtxt(load_path+'real_world_samples_input.csv', delimiter=',')
Y = np.loadtxt(load_path+'real_world_samples_output.csv', delimiter=',')
X2 = np.r_[X, np.loadtxt(load_path+"policy"+str(loop_num)+'/simulation2/simulation_samples_input.csv' , delimiter=',')]
Y2 = np.r_[Y, np.loadtxt(load_path+"policy"+str(loop_num)+'/simulation2/simulation_samples_output.csv', delimiter=',')]
print("prev_log_evidence =",pm_class.log_evidence())
print("debug_log_evidence =",pm_class.log_evidence_of_new_data(X,Y))
print("new_log_evidence =",pm_class.log_evidence_of_new_data(X2,Y2))
temp_data = np.array([pm_class.log_evidence(), pm_class.log_evidence_of_new_data(X2,Y2)])
np.savetxt(save_path+"policy"+str(loop_num)+'/simulation2/log_evidence.csv',temp_data,delimiter=',')
np.savetxt(save_path+"policy"+str(loop_num)+'/log_evidence_for_learn2.csv',np.array([pm_class.log_evidence()]),delimiter=',', fmt ='%.6f')
return temp_data[1]
def custom_pendulum_wrap(
env,
model_switch=0,
dirname="./data_debug/",
filename1="debug_input.csv",
filename2="debug_output.csv",
):
# 1 #
if model_switch < 9:
real_dynamics = pendulum_real_model.PendulumDynamics()
else:
real_dynamics = pendulum_real_model_pilco.PendulumDynamics()
real_dynamics.logger_parameter()
real_dynamics.wrap_env(env)
# 2 #
if model_switch > 0:
dataX = np.loadtxt(dirname + filename1, delimiter=',')
dataY = np.loadtxt(dirname + filename2, delimiter=',')
if model_switch < 9:
init_s = np.loadtxt(dirname + 'current_state.csv', delimiter=',')
else:
init_s = None
if 1 == model_switch or 3 == model_switch or 4 == model_switch or 11 == model_switch or 111 == model_switch:
npm_class = pendulum_npm.PendulumNPM(dataX, dataY, init_state=init_s)
npm_class.logger_parameter()
if 2 == model_switch or 3 == model_switch or 4 == model_switch:
pm_class = pendulum_pm.PendulumPM(dataX, dataY, init_state=init_s)
pm_class.logger_parameter()
# 3 #
best_fit_model = 0
if 1 == model_switch or 2 == model_switch or 5 == model_switch:
best_fit_model = model_switch
if 3 == model_switch:
log_evidence = pm_class.log_evidence() - npm_class.log_evidence()
if log_evidence > 0.:
best_fit_model = 2
else:
best_fit_model = 1
logger.log("log_evidence_diff =", log_evidence)
if 4 == model_switch:
npmcv = npm_class.k_fold_cv(dataX, dataY, k=10)
pmcv = pm_class.k_fold_cv(dataX, dataY, k=10)
cv = pmcv - npmcv
if cv > 0.:
best_fit_model = 2
else:
best_fit_model = 1
# 4 #
if 1 == best_fit_model:
npm_class.wrap_env(env)
if 2 == best_fit_model:
pm_class.wrap_env(env)
if 11 == model_switch or 111 == model_switch: # only for comparison with pilco
npm_class.thdot_clip_value = 50. # only for comparison with pilco
# In this case, the resulting real-world velocity exceeds the max_speed given for accelerating learning process for main results.
npm_class.wrap_env_pilco(env) # only for comparison with pilco
npm_class.logger_parameter()
best_fit_model = model_switch
logger.log("flag in custom_pendulum_wrap =", model_switch)
logger.log("actually_selected_model=", best_fit_model)
if (best_fit_model + model_switch) > 0:
logger.log("init_state=", init_s)
return best_fit_model
real_dynamics.wrap_env(env.env)
dn ="./data_debug2/"
fn1="debug_input.csv"
fn2="debug_output.csv"
#pendulum_real_model.generate_test_samples(dirname=dn,filename1=fn1,filename2=fn2)
dataX = np.loadtxt(dn+fn1, delimiter=',')
dataY = np.loadtxt(dn+fn2, delimiter=',')
test_class = PendulumPM(dataX,dataY)
test_class.wrap_env(env.env)
test_class.logger_parameter()
import pendulum_pm # add
test2_class = pendulum_pm.PendulumPM(dataX,dataY) # add
test2_class.wrap_env(env.env) # add
test2_class.logger_parameter() # add
''' # add
episode_count=0
while episode_count<1:
env.reset()
env.env.state= np.array([[np.pi],[0.]])
while True:
ac = env.action_space.sample()
ob, rew, new, _ = env.step(ac)
env.render()
if new:
episode_count +=1