def noescExpTest():
env = ('AttFC_GyroErr-MotorVel_M4_Ep-v0',
'AttFC_GyroErr-MotorVel_M4_Con-v0', 'AttFC_GyroErr1_M4_Ep-v0')
cur_env = env[2]
tp = tg.trainParams()
pp = tg.policyParams()
# Set up parameters
tp.timesteps_per_actorbatch = 500
tp.optim_batchsize = 32
tp.optim_epochs = 5
# Set training length
num_eps = 1
tp.num_timesteps = num_eps * 1000
# Name Model
tp.modelName('z10')
#Run Training
with tg.U.tf.Graph().as_default():
tg.train(tp, pp, cur_env)
# Model Evaluation
me = eg.ModelEval(tp.model_name, cur_env)
me.evalModel(15)
me.saveEval()
def runEvals(exps):
#FINISH
# Run and save an evaluation for all trained models
for i in range(len(exps['Model'])):
me = eg.ModelEval(exps['Model'][i], 'AttFC_GyroErr1_M4_Ep-v0')
me.evalModel(50)
me.saveEval()
def checkR(model):
me = eg.ModelEval(model, 'AttFC_GyroErr1_M4_Ep-v0')
me.evalModel(3)
print('Average Reward: ', me.proc_eval.eps_avg_r)
print('Max Rewards: ', me.proc_eval.eps_max_r)
print('Rise times: ')
print(me.proc_eval.eps_r_rise,
me.proc_eval.eps_p_rise,
me.proc_eval.eps_y_rise,
sep='\n')
print('-----Beginning-----')
print('Rewards: ',
me.eps[0]['rewards'][0:4],
me.eps[1]['rewards'][0:4],
me.eps[2]['rewards'][0:4],
sep='\n')
print('-----End-----')
print('Rewards: ',
me.eps[0]['rewards'][990:994],
me.eps[1]['rewards'][990:994],
me.eps[2]['rewards'][990:994],
sep='\n')
print('-------------')
print('Desired: ', [
me.eps[0]['droll_v'][0], me.eps[0]['dpitch_v'][0],
me.eps[0]['dyaw_v'][0]
])
print('Total R: ', np.sum(me.eps[0]['rewards']),
np.sum(me.eps[1]['rewards']), np.sum(me.eps[2]['rewards']))
def initRewards(exps):
# FINISH
# Creates rewards from untrained models to compare to trained models to assess reward elements learned
me = eg.ModelEval(exps['Model'][i], 'AttFC_GyroErr1_M4_Ep-v0')
me.evalModel()
me.saveEval(save_name='init' + str(i))