#read in the weights as a 2-d array and the feature counts of the policy
W = helper.get_weightchain_array("../../mcmc_data/" + args.env_name + "_0.txt")
print(np.mean(W, axis=0))
eval_policies = ['00025', '00325', '00800', '01450']
if args.env_name == "enduro":
eval_policies = ['03125', '03425', '03900', '04875']
gt_return_list = []
fcount_list = []
return_dist_list = []
print(" policy & mean & 0.05-VaR & ave length & min & stdev")
for eval in eval_policies:
#print("-"*20)
#print("eval", eval)
returns, fcounts = helper.parse_avefcount_array('../../policies/' +
args.env_name + '_' +
eval + '_fcounts_100.txt')
return_dist = np.dot(W, fcounts)
print(
"{} & {:.1f} & {:.1f} & {:.1f} & {:.1f} & {:.0f} & {:.0f} \\\\".format(
eval, np.mean(return_dist),
helper.worst_percentile(return_dist, 0.01), fcounts[-1],
np.mean(returns), np.min(returns), np.std(returns)))
gt_return_list.append(returns)
fcount_list.append(fcounts)
return_dist_list.append(return_dist)
plt.figure(0)
plt.hist(return_dist_list,
30,
#read in the weights as a 2-d array and the feature counts of the policy
W, log_lik = helper.get_weightchain_array(args.mcmc_file, return_likelihood=True)
print(np.mean(W, axis=0))
eval_policies = ['00025', '00325', '00800', '01450', 'mean', 'map']
name_transform = {'00025':'policy A', '00325':'policy B', '00800':'policy C', '01450':'policy D', 'mean':'mean', 'map':'MAP', 'noop': 'no-op'}
if args.env_name == "enduro":
eval_policies =['03125', '03425', '03900', '04875', 'mean', 'map']
name_transform = {'03125':'policy A', '03425':'policy B', '03900':'policy C', '04875':'policy D', 'mean':'mean', 'map':'MAP', 'noop': 'no-op'}
gt_return_list = []
fcount_list = []
return_dist_list = []
print(" policy & mean & " + str(alpha) + "-VaR & ave length & gt & min gt \\\\ \hline")
for eval in eval_policies:
#print("-"*20)
#print("eval", name_transform[eval])
returns, fcounts = helper.parse_avefcount_array('../../policies/' + args.env_name +'_' + eval + '_fcounts_onehot_truncated_terminal' + str(args.no_term) + '.txt')
#print(fcounts)
return_dist = np.dot(W,fcounts)
print("{} & {:.1f} & {:.1f} & {:.1f} & {:.1f} & {:.0f} \\\\".format(name_transform[eval], np.mean(return_dist), helper.worst_percentile(return_dist, alpha), np.sum(fcounts), np.mean(returns), np.min(returns)))
gt_return_list.append(returns)
fcount_list.append(fcounts)
return_dist_list.append(return_dist)
if args.env_name == "breakout":
#I realized that I need to rerun the noop code for the full features. I keep overwriting it.
#evaluate the no-op policy
#returns, fcounts = helper.parse_avefcount_array('../../policies/breakout_noop_fcounts.txt')
returns = np.array([0,0])
if args.no_term:
params = {
'legend.fontsize': 'x-large',
'figure.figsize': (5, 4),
'axes.labelsize': 'x-large',
'axes.titlesize': 'x-large',
'xtick.labelsize': 'x-large',
'ytick.labelsize': 'x-large'
}
plt.rcParams.update(params)
#plt.style.use('seaborn-deep')
parser = argparse.ArgumentParser(description=None)
#parser.add_argument('--env_name', help="name of the environment, e.g. 'breakout'")
args = parser.parse_args()
eval_games = ['beamrider', 'breakout', 'enduro', 'seaquest', 'spaceinvaders']
gt_return_list = []
fcount_list = []
return_dist_list = []
for game in eval_games:
print(game)
for eval in ["mean", "map"]:
print(eval)
#print("eval", eval)
returns, fcounts = helper.parse_avefcount_array('../../policies/' +
game + '_' + eval +
'_fcounts.txt')
print("{:.0f} & {:.1f}".format(np.min(returns), np.mean(returns)))
#read in the weights as a 2-d array and the feature counts of the policy
W = helper.get_weightchain_array("../../mcmc_data/" + args.env_name + "_gt_chain.txt")
print(np.mean(W, axis=0))
eval_policies = ['00025', '00325', '00800', '01450', 'mean', 'map']
name_transform = {'00025':'policy A', '00325':'policy B', '00800':'policy C', '01450':'policy D', 'mean':'mean', 'map':'MAP', 'noop': 'no-op'}
if args.env_name == "enduro":
eval_policies =['03125', '03425', '03900', '04875', 'mean', 'map']
name_transform = {'03125':'policy A', '03425':'policy B', '03900':'policy C', '04875':'policy D', 'mean':'mean', 'map':'MAP', 'noop': 'no-op'}
gt_return_list = []
fcount_list = []
return_dist_list = []
print(" policy & mean & 0.05-VaR & ave length & gt & min gt \\\\ \hline")
for eval in eval_policies:
#print("-"*20)
#print("eval", eval)
returns, fcounts = helper.parse_avefcount_array('../../policies/' + args.env_name +'_' + eval + '_fcounts_gt.txt')
return_dist = np.dot(W,fcounts)
print("{} & {:.1f} & {:.1f} & {:.1f} & {:.1f} & {:.0f} \\\\".format(name_transform[eval], np.mean(return_dist), helper.worst_percentile(return_dist, alpha), fcounts[-1], np.mean(returns), np.min(returns)))
gt_return_list.append(returns)
fcount_list.append(fcounts)
return_dist_list.append(return_dist)
if args.env_name == "breakout":
#evaluate the no-op policy
returns, fcounts = helper.parse_avefcount_array('../../policies/breakout_noop_fcounts.txt')
#normalize
return_dist = np.dot(W,fcounts)
print("{} & {:.1f} & {:.1f} & {:.1f} & {:.1f} & {:.0f} ".format("no-op", np.mean(return_dist), helper.worst_percentile(return_dist, 0.05), fcounts[-1], np.mean(returns), np.min(returns)))
plt.show()
eval_policies = ['00025', '00325', '00800', '01450', 'mean', 'map']
name_transform = {'00025':'policy A', '00325':'policy B', '00800':'policy C', '01450':'policy D', 'mean':'mean', 'map':'MAP', 'noop': 'no-op'}
if args.env_name == "enduro":
eval_policies =['03125', '03425', '03900', '04875', 'mean', 'map']
name_transform = {'03125':'policy A', '03425':'policy B', '03900':'policy C', '04875':'policy D', 'mean':'mean', 'map':'MAP', 'noop': 'no-op'}
gt_return_list = []
fcount_list = []
return_dist_list = []
print(" policy & mean & " + str(args.alpha) + "-VaR & mu+10*Var & ave length & gt & min gt \\\\ \hline")
for eval in eval_policies:
#print("-"*20)
#print("eval", eval)
returns, fcounts = helper.parse_avefcount_array('../../policies/' + args.env_name +'_' + eval + args.identifier + '.params_stripped.params_fcounts_auxiliary.txt')
#print("num rollouts", len(returns))
return_dist = np.dot(W,fcounts)
print("{} & {:.1f} & {:.1f} & {:.1f} & {:.1f} & {:.1f} & {:.0f} \\\\".format(name_transform[eval], np.mean(return_dist), helper.worst_percentile(return_dist, args.alpha), 10*helper.worst_percentile(return_dist, args.alpha) + np.mean(return_dist), 0, np.mean(returns), np.min(returns)))
gt_return_list.append(returns)
fcount_list.append(fcounts)
return_dist_list.append(return_dist)
if args.env_name == "breakout" and args.noop:
#evaluate the no-op policy
returns, fcounts = helper.parse_avefcount_array('../../policies/breakout_no-op' + args.identifier + '.params_stripped.params_fcounts_auxiliary.txt')
noop_returns = returns
#normalize
return_dist = np.dot(W,fcounts)