def plot_env_cg():
'''Experiments with varying Gopt'''
env = envs[0]
plot_rwds(cg_labels, cdict=color_dict, fdir=fdir + env + '-v1/', model_paths=model_paths, BEST=BEST,
MAX_LEN=MAX_LEN(env), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title='Gradient Optimizers')
return
def plot_env_varnoise(env, fnoise_dir='../new_final_results/Swimmer-v1/NOISE/', noise_vals=np.arange(0.0, 1.0, 0.1)):
''' Experiments with obsnoise'''
noise_paths = {}
[noise_paths.update({'joint_%.1f' % noise: '%.1f/lite-cont/lite-cont.pkl' % noise,
'obs_%.1f' % noise: '%.1f/obsVR/obsVR.pkl' % noise,
'AR_%.1f' % noise: '%.1f/arVR/arVR.pkl' % noise}) for noise in noise_vals]
colors_noise1 = plt.cm.winter(np.linspace(0, 1, len(noise_vals)))
colors_noise2 = plt.cm.autumn(np.linspace(0, 1, len(noise_vals)))
colors_noise3 = plt.cm.cool(np.linspace(0, 1, len(noise_vals)))
ncolor_dict = {}
ncolor_dict.update(dict(
[(k, colors_noise1[-1]) for (i, k) in enumerate(list(filter(lambda k: k[:4] == 'join', noise_paths.keys())))]))
ncolor_dict.update(dict(
[(k, colors_noise2[-1]) for (i, k) in enumerate(list(filter(lambda k: k[:3] == 'obs', noise_paths.keys())))]))
ncolor_dict.update(dict(
[(k, colors_noise3[-1]) for (i, k) in enumerate(list(filter(lambda k: k[:2] == 'AR', noise_paths.keys())))]))
cn_dict = lambda key: ncolor_dict.get(key)
n_paths = lambda key: noise_paths.get(key)
nkeys = [['joint_%.1f' % noise, 'obs_%.1f' % noise, 'AR_%.1f' % noise] for noise in noise_vals];
for i, i_nkeys in enumerate(nkeys):
plot_rwds(i_nkeys, cdict=cn_dict, fdir=fnoise_dir,
model_paths=n_paths, BEST=BEST, MAX_LEN=MAX_LEN(env), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title='Noise_performance%.1f' % noise_vals[i])
return
def plot_env_opt():
'''Experiments for RPSP opts'''
env = envs[0]
plot_rwds(opt_labels, cdict=color_dict, fdir=fdir + env + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(env), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title='RPSP optimizers')
return
def plot_envs_vrpggru():
'''Experiments for RPSP/gru minimal comparison'''
for i in range(len(envs)):
plot_rwds(VRPGgru_labels, cdict=color_dict, fdir=fdir + envs[i] + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title=envs[i] + ' VRPGvsGRU comparison', step=max_lims[envs[i]][1], shape_dict=shape_dict2)
return
def plot_env_init():
'''Experiments with random initialization'''
env = envs[0]
plot_rwds(init_labels, cdict=color_dict, fdir=fdir + env + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(env), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title='RPSP Initialization')
return
def plot_envs_ar():
'''Experiments with varying window'''
for i in range(len(envs)):
plot_rwds(AR_labels, cdict=color_dict, fdir=fdir + envs[i] + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title=envs[i] + ' FM $R_{iter}$ ', step=max_lims[envs[i]][1], shape_dict=shape_dict,
kfold=max_lims[envs[i]][1] * MAX_TRIALS)
return
def plot_envs_rpsp():
'''Experiments for RPSP variants'''
for i in range(len(envs)):
plot_rwds(RPSP_labels, cdict=color_dict, fdir=fdir + envs[i] + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title=envs[i] + ' RPSP variants $R_{iter}$ ', step=max_lims[envs[i]][1], shape_dict=shape_dict,
kfold=max_lims[envs[i]][1] * MAX_TRIALS)
return
def plot_envs_gru():
for i in range(len(envs)):
'''plot for GRUs'''
plot_rwds(gru_labels, cdict=color_dict, fdir=fdir + envs[i] + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title=envs[i] + r' GRU $R_{iter}$ ', step=max_lims[envs[i]][1], shape_dict=shape_dict,
kfold=max_lims[envs[i]][1] * MAX_TRIALS)
return
def plot_envs_reactive():
''' Experiments for each environment'''
for i in range(len(envs)):
env = envs[i]
plot_rwds(reactive_labels, cdict=color_dict, fdir=fdir + envs[i] + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title=envs[i] + 'Prediction as regularization $R_{iter}$ ', shape_dict=shape_dict3, ncol=3,
kfold=max_lims[envs[i]][1] * MAX_TRIALS)
return
def plot_envs_rpspgru():
'''Experiments for RPSP/gru minimal comparison'''
for i in range(len(envs)):
env = envs[i]
plot_rwds(RPSPgru_labels, cdict=best_colors(envs[i]), fdir=fdir + envs[i] + '-v1/',
model_paths=best_paths(envs[i]), BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP,
MAX_TRIALS=MAX_TRIALS,
title=envs[i] + ' GRU Filter comparison $R_{iter}$ ', step=max_lims[envs[i]][1],
shape_dict=best_shape(envs[i]),
kfold=max_lims[envs[i]][1] * MAX_TRIALS)
return
def plot_envs_overall():
"""
plot overall environment performance
@return:
"""
AUC = []
R_iter = []
for i in range(len(envs)):
env = envs[i]
print(env)
bestmodels, R_cum, aucurve = plot_rwds(envs_labels, cdict=color_dict,
fdir=fdir + envs[i] + '-v1/',
model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(envs[i]),
STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title=envs[i] + ' $R_{iter}$ ', step=max_lims[envs[i]][1],
shape_dict=shape_dict,
kfold=max_lims[envs[i]][1] * MAX_TRIALS)
print('done print')
R_iter.append(bestmodels)
R = np.asarray(aucurve, dtype=float)
AUC.extend(zip(np.mean(R, axis=1), np.std(R, axis=1) / np.sqrt(R.shape[1])))
print('done. Computing AUC')
for j in [1, 2]:
for m in range(len(envs_labels) - 1):
print_ttest(envs_labels, R, m, j)
print('\n')
# Swimmer
# t-test ( FM RPSP-VRPG )= [ -0.676586869332 , 0.507653821772 ]
# t-test ( GRU RPSP-VRPG )= [ -11.4782215848 , 9.9484555701e-07 ]
# t-test ( RPSP-Alt RPSP-VRPG )= [ -3.10454361915 , 0.00933872173406 ]
# t-test ( FM RPSP-Alt )= [ 1.73508240067 , 0.111134237192 ]
# t-test ( GRU RPSP-Alt )= [ -20.2985459977 , 2.4045809148e-09 ]
# Hopper
# t-test ( FM RPSP-VRPG )= [ -1.97496384693 , 0.0638921448438 ]
# t-test ( GRU RPSP-VRPG )= [ -2.98367426585 , 0.0121047687853 ]
# t-test ( RPSP-Alt RPSP-VRPG )= [ 2.25405603636 , 0.0382107798284 ]
# t-test ( FM RPSP-Alt )= [ -3.92498523554 , 0.00123896806466 ]
# t-test ( GRU RPSP-Alt )= [ -4.909161505 , 0.000577310309778 ]
# Walker2d
# t-test ( FM RPSP-VRPG )= [ -0.0793340586349 , 0.937667000948 ]
# t-test ( GRU RPSP-VRPG )= [ -2.9657440937 , 0.00870691843556 ]
# t-test ( RPSP-Alt RPSP-VRPG )= [ 2.23688724131 , 0.0412540309747 ]
# t-test ( FM RPSP-Alt )= [ -2.05494189388 , 0.0601025184967 ]
# t-test ( GRU RPSP-Alt )= [ -6.51606587482 , 5.3677973935e-06 ]
# CartPole
# t-test ( FM RPSP-VRPG )= [ -4.49013758712 , 0.00103201257316 ]
# t-test ( GRU RPSP-VRPG )= [ 1.23597901022 , 0.232353991956 ]
# t-test ( RPSP-Alt RPSP-VRPG )= [ -0.358340067592 , 0.726097854222 ]
# t-test ( FM RPSP-Alt )= [ -8.06880999537 , 6.44596048562e-07 ]
# t-test ( GRU RPSP-Alt )= [ 1.95090137725 , 0.0740240842398 ]
# print (np.round(np.asarray(AUC)[[0, 1, 2, 3, 5, 9], :] / 1000.0, decimals=1))
# print (reg_labels, filter_labels)
return R_iter, AUC
def plot_envs_reg():
'''Experiments for RPSP/gru minimal comparison'''
auc_filter = []
for i in range(len(envs)):
env = envs[i]
x, x, aucurve = plot_rwds(reg_labels, cdict=best_colors(envs[i]), fdir=fdir + envs[i] + '-v1/',
model_paths=best_paths(envs[i]), BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP,
MAX_TRIALS=MAX_TRIALS,
title=envs[i] + ' Filter regularization $R_{iter}$ ', step=max_lims[envs[i]][1],
shape_dict=best_shape(envs[i]), ylim=[-10, max_lims[envs[i]][0]],
kfold=max_lims[envs[i]][1] * MAX_TRIALS)
Rfilter = np.asarray(aucurve, dtype=float)
auc_filter.extend(zip(np.mean(Rfilter, axis=1), np.std(Rfilter, axis=1) / np.sqrt(Rfilter.shape[1])))
for k in [0, 1, 3]:
print_ttest(reg_labels, Rfilter, 2, k)
return auc_filter
def plot_envs_filter():
'''Experiments for RPSP/gru minimal comparison'''
gru_auc = []
for i in range(len(envs)):
env = envs[i]
x, x, aucurve = plot_rwds(filter_labels, cdict=best_colors(envs[i]),
fdir=fdir + envs[i] + '-v1/',
model_paths=best_paths(envs[i]), BEST=BEST, MAX_LEN=MAX_LEN(envs[i]), STEP=STEP,
MAX_TRIALS=MAX_TRIALS,
title=envs[i] + ' PSR vs. GRU filtering $R_{iter}$ ', step=max_lims[envs[i]][1],
shape_dict=best_shape(envs[i]), ylim=[-10, max_lims[envs[i]][0]],
kfold=max_lims[envs[i]][1] * MAX_TRIALS, ncol=5)
Rgru = np.asarray(aucurve, dtype=float)
print_ttest(filter_labels, Rgru, 1, 0)
gru_auc.extend(zip(np.mean(Rgru, axis=1), np.std(Rgru, axis=1) / np.sqrt(Rgru.shape[1])))
return gru_auc
def plot_envs_noise():
for i in range(len(envs)):
env = envs[i]
R_cumulative = []
f, ax = plt.subplots(nrows=1, ncols=len(noise_levels), figsize=(15, 6), sharey=True)
ax[0].set_ylabel(envs[i] + ' $R_{iter}$ ')
for j, noise in enumerate(noise_levels):
R_iter, R_cum, auc = plot_rwds(noise_labels, cdict=color_dict, fdir=fdir + envs[i] + '-v1/',
model_paths=noise_paths(noise), BEST=BEST, MAX_LEN=MAX_LEN(envs[i]),
STEP=STEP, MAX_TRIALS=MAX_TRIALS,
shape_dict=shape_dict3, ylim=[-10, 800], # max_lims[envs[i]][0]],
f=f, ax=ax[j], legend=(j == 1), kfold=max_lims[envs[i]][1] * MAX_TRIALS)
# title=envs[i]+' Observation noise ',
R_cumulative.extend(R_cum)
plt.subplots_adjust(wspace=0.1, top=0.9, bottom=0.2, left=0.08, right=0.99, )
ax[j].set_title("noise $\sigma$=" + str(noise), position=(0.5, 0.92), fontsize=20)
f.savefig(fdir + envs[i] + '-v1/' + envs[i] + 'noiseobs_rwdperiter_best%d_step%d.png' % (BEST, STEP),
dpi=300)
# print 'donenoise'
# plot_table(R_cumulative, noise_labels, noise_levels, envs[i]+' Obstacle noise AUC', 'AUC (cumulative Return $10^3$)', fdir+env+'-v1/', incr=1000)
return
def plot_env_ar():
'''Experiments with varying window'''
env = envs[0]
plot_rwds(AR_labels, cdict=color_dict, fdir=fdir + env + '-v1/', model_paths=model_paths, BEST=BEST, MAX_LEN=MAX_LEN(env), STEP=STEP, MAX_TRIALS=MAX_TRIALS,
title='AR window')
return