def _collect_aspect_highlights(_aspect_name, aspect_lst, num_highlights):
global aspect_group, highlights_dirs, ag_idx
highlights = []
aspect_idx = 0
total_highlights = 0
highlight_idx = 0
# until highlight budget is not reached
group_name = aspect_group[_aspect_name]
while total_highlights < num_highlights and aspect_idx < len(aspect_lst):
# tries to load highlight for the aspect
s, *_ = aspect_lst[aspect_idx]
highlight_file = join(
highlights_dirs[ag_idx], group_name,
'{}-{}-video{}.{}'.format(_aspect_name, s, highlight_idx,
NUMPY_EXTENSION))
if exists(highlight_file):
highlights.append(np.load(highlight_file))
total_highlights += 1
highlight_idx += 1
print_line('Loaded highlight from: {}'.format(highlight_file),
log_file)
else:
# advances aspect element
aspect_idx += 1
highlight_idx = 0
return highlights
def save_stats(self, path, clear=True, img_format='pdf'):
super().save_stats(path, clear)
# collects and prints final stats to file
e = self.config.num_episodes
with open(join(path, 'results.log'), 'w') as file:
print_line('\nStats (avg. of {} episodes):'.format(e), file)
var_names = list(self.stats_collector.all_variables())
var_names.sort()
self._print_stats(e, var_names, file)
def _print_stats(self, e, selected_var_names, file=None):
"""
Prints the current stats of a selected number of variables to the screen.
:param int e: the number of the episode.
:param list selected_var_names: a list of the names of the variables for which to print the statistics.
:param stream file: the file on which to save the message line.
:return:
"""
for var_name in selected_var_names:
avg, std = self.stats_collector.get_mean(var_name, e)
print_line(
'Avg {}: {:.2f} ± {:.2f}'.format(var_name.lower(), avg, std),
file)
def _save_highlights_for_aspect(sub_dir, _aspect_name, aspect_lst,
num_highlights, agent_name):
global log_file
print_line('==============================================', log_file)
print_line(
'Collecting and saving highlights for aspect {} of \'{}\' in sub-dir: \'{}\''
.format(_aspect_name, agent_name, sub_dir), log_file)
# collect and saves highlights for this aspect
highlights = _collect_aspect_highlights(_aspect_name, aspect_lst,
num_highlights)
_save_highlights_video(highlights, sub_dir, agent_name)
def print_variable_evo_comparison(_output_dir, _stats, _var_name):
print_line('\n==============================================', log_file)
print_line(
'Processing comparison of evolution for \'{}\'...\n'.format(
_var_name.lower()), log_file)
# plots and creates data frame from trial data for each variable
plt.figure()
ag_names = list(_stats.keys())
color_map = plt.cm.get_cmap(COLOR_SET_NAME)(np.arange(len(ag_names)))
var_data = OrderedDict()
for _ag_idx, _ag_name in enumerate(ag_names):
means, stds, counts = _stats[_ag_name][_var_name]
var_data['{} Mean'.format(_ag_name)] = means
var_data['{} Std'.format(_ag_name)] = stds
var_data['{} Count'.format(_ag_name)] = counts
errs = stds / np.sqrt(counts)
plt.plot(smooth_data(means), label=_ag_name, c=color_map[_ag_idx])
plt.fill_between(np.arange(len(means)),
smooth_data(means - errs),
smooth_data(means + errs),
color=color_map[_ag_idx],
alpha=0.5)
df = pd.DataFrame(var_data)
df.to_csv(join(subdir, '{}-trial-data.csv'.format(_var_name.lower())))
# gets y label adjust for base 10
exp = get_y_label_adjust_base(df.max().max())
y_label = _var_name.title()
# configures axes and legend
plt.legend(fontsize=16)
plt.xlabel('Episode', fontweight='bold', fontsize=14)
if abs(exp) >= 2:
y_label = '{} $\\mathregular{{\\left(\\times10^{}\\right)}}$'.format(
y_label, exp)
plt.gca().yaxis.set_major_formatter(
FuncFormatter(lambda x, _: '{}'.format(int(x / (10**exp)))))
plt.ylabel(y_label, fontweight='bold', fontsize=14)
plt.xlim(0, df.shape[0])
# saves figure
plt.savefig(join(_output_dir, 'evo {}.pdf'.format(_var_name.lower())),
pad_inches=.02,
bbox_inches='tight')
plt.close()
def _print_actions(title, diff_actions, table, file, write_console):
print_line('\t\t{}:\t['.format(title), file, write_console, False)
for a in range(len(table)):
print_line(
str(round(table[a], 3)).rjust(10), file, write_console, False)
print_line('*' if a in diff_actions else ' ', file, write_console,
False)
print_line(']\n', file, write_console, False)
def _save_highlights_for_aspects(sub_dir, _aspect_names, aspect_lsts,
num_highlights, agent_name):
global log_file
print_line('==============================================', log_file)
print_line(
'Collecting and saving highlights for {} aspects of \'{}\' in sub-dir: \'{}\'...'
.format(len(_aspect_names), agent_name, sub_dir), log_file)
# collect and saves highlights for the different aspects
highlights = []
for _i in range(len(_aspect_names)):
highlights.extend(
_collect_aspect_highlights(_aspect_names[_i], aspect_lsts[_i],
num_highlights[_i]))
_save_highlights_video(highlights, sub_dir, agent_name)
def _save_highlights_video(highlights, sub_dir_name, _file_name):
global log_file, output_dir
sub_dir = join(output_dir, sub_dir_name)
if not exists(sub_dir):
makedirs(sub_dir)
# tries to save highlight compact to video file
if len(highlights) > 0:
highlights_frames = []
for frames in highlights:
highlights_frames.extend(frames[frames.files[0]])
video_file_path = join(
sub_dir, '{}.{}'.format(_file_name.lower(), VIDEO_EXTENSION))
record_video(highlights_frames, video_file_path, HIGHLIGHTS_FPS)
print_line('Saved highlights compact in: {}'.format(video_file_path),
log_file)
def _save_sequence(sub_dir_name, sequence_lst, agent_name):
global output_dir, log_file, ag_idx
sub_dir = join(output_dir, sub_dir_name)
if not exists(sub_dir):
makedirs(sub_dir)
print_line('==============================================', log_file)
print_line(
'Collecting and saving sequence for \'{}\' in sub-dir: \'{}\''.format(
agent_name, sub_dir), log_file)
# tries to get the longest recorded sequences
for seq_idx, seq_info in enumerate(sequence_lst):
seq_name, _ = SequencesExplainer.get_sequence_name(seq_idx, seq_info)
seq_file_name = join(
sequences_dirs[ag_idx],
SequencesExplainer.get_sequence_file_name(seq_name, 0))
# if file exists, copy to results dir
if exists(seq_file_name):
video_file_path = join(
sub_dir, '{}.{}'.format(agent_name.lower(), VIDEO_EXTENSION))
shutil.copyfile(seq_file_name, video_file_path)
print_line(
'Saved sequence by copying from: {}\n\tto: {}'.format(
seq_file_name, video_file_path), log_file)
return
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
action_names = self.config.get_action_names()
print_line('====================================', file, write_console)
print_line(
'{} certain sequences to sub-goals found (min. prob > 0):'.format(
len(self.certain_seqs_to_subgoal)), file, write_console)
for s, n, seq, prob in self.certain_seqs_to_subgoal:
self._print_sequence(s, n, seq, prob, feats_nbins, action_names,
file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} states with an uncertain future found:'.format(
len(self.uncertain_future_states)), file, write_console)
for s, n in self.uncertain_future_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{} (count: {})'.format(feats_labels, n), file,
write_console)
def print_variable_mean_comparison(_output_dir, _stats, _var_name):
print_line('\n==============================================', log_file)
print_line(
'Processing comparison of means for \'{}\'...\n'.format(
_var_name.lower()), log_file)
# reorganizes data by stats x agent
ag_names = list(_stats.keys())
var_data = np.zeros((3, len(_stats)))
for _ag_idx, _ag_name in enumerate(ag_names):
means, stds, counts = _stats[_ag_name][_var_name]
mean, std, count = get_combined_mean(means, stds, counts)
var_data[:, _ag_idx] = [mean, std, count]
# convert variable stats to pandas data-frame
df = pd.DataFrame(var_data,
index=['Mean', 'Std', 'Count'],
columns=ag_names)
# saves to CSV
file_name = _var_name.lower()
df.to_csv(join(_output_dir, '{}.csv'.format(file_name)))
# print to screen
print_line(str(df), log_file)
x_axis = np.arange(len(ag_names))
means = [var_data[0][i] for i in x_axis]
errors = [var_data[1][i] / np.sqrt(var_data[2][i]) for i in x_axis]
# errors = [var_data[1][i] for i in x_axis]
color_map = plt.cm.get_cmap(COLOR_SET_NAME)(x_axis)
# gets y label adjust for base 10
exp = get_y_label_adjust_base(max(means) + max(errors))
y_label = 'Mean {}'.format(_var_name).title()
# print bar chart of means to image file (may not work well for all variables)
plt.bar(x_axis,
means,
yerr=errors,
capsize=10,
align='center',
color=color_map)
plt.xticks(x_axis, ag_names)
plt.xlabel('Agent', fontweight='bold', fontsize=14)
if abs(exp) >= 2:
y_label = '{} $\\mathregular{{\\left(\\times10^{}\\right)}}$'.format(
y_label, exp)
plt.gca().yaxis.set_major_formatter(
FuncFormatter(lambda x, _: '{}'.format(int(x / (10**exp)))))
plt.ylabel(y_label, fontweight='bold', fontsize=14)
plt.savefig(join(_output_dir, '{}.pdf'.format(file_name)),
pad_inches=.02,
bbox_inches='tight')
plt.close()
def _print_sequence(self, s, n, seq, prob, feats_nbins, action_names, file,
write_console):
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line(
'\t{}-{} (count: {}, prob. reaching target: {:.3e})'.format(
s, feats_labels, n, prob), file, write_console)
for a, ns in seq:
obs_vec1 = get_features_from_index(ns, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec1, True)
obs_vec_diff = self._vec_diff(obs_vec1, obs_vec)
feats_labels = [feats_labels[i] for i in obs_vec_diff]
obs_vec = obs_vec1
print_line(
'\t\t{} -> {}-{}'.format(action_names[a], ns, feats_labels),
file, write_console)
print_line('____________________________________', file, write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
print_line('====================================', file, write_console)
print_line('Total states visited: {}'.format(self.total_count), file,
write_console)
print_line(
'Coverage of the state-space: {:.2f}%'.format(
self.state_coverage * 100.), file, write_console)
print_line(
'Dispersion of the visits to the state-space: {:.3f}'.format(
self.state_dispersion), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} frequent states found (min. support: {}):'.format(
len(self.freq_states), self.min_state_count), file,
write_console)
for s, n in self.freq_states:
rf = n / float(self.total_count)
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line(
'\t{}-{} (count: {}, freq.: {:.3f})'.format(
s, feats_labels, n, rf), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} infrequent states found (max. support: {}):'.format(
len(self.infreq_states), self.max_state_count), file,
write_console)
for s, n in self.infreq_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} (count: {})'.format(s, feats_labels, n), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} frequent feature-sets found (min. support: {}, min. assoc. strength: {}):'
.format(len(self.freq_feature_sets), self.min_feat_set_count,
self.min_feat_set_assoc), file, write_console)
num_feats = len(feats_nbins)
for feat_set, jacc in self.freq_feature_sets:
feats_labels = self._get_feats_labels(feat_set, num_feats)
print_line('\t{} (jacc: {:.3f})'.format(feats_labels, jacc), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} interesting feature-rules found (min. confidence: {}):'.format(
len(self.freq_feature_rules), self.min_feat_rule_conf), file,
write_console)
for ant, cons, n, conf in self.freq_feature_rules:
n /= float(self.total_count)
ant = self._get_feats_labels(ant, num_feats)
cons = self._get_feats_labels(cons, num_feats)
print_line(
'\t{} => {} (freq: {:.3f}, conf: ({:.3f}))'.format(
ant, cons, n, conf), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} infrequent feature-sets found (max. assoc. strength: {}):'.
format(len(self.infreq_feature_sets),
self.max_feat_set_assoc), file, write_console)
for feat_set in self.infreq_feature_sets:
feats_labels = self._get_feats_labels(feat_set, num_feats)
print_line('\t{}'.format(feats_labels), file, write_console)
def _save_report(self, file, write_console):
print_line('\n===================================================================', file, write_console)
print_line('TRANSITION ANALYSIS', file, write_console)
self.trans_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('REWARD ANALYSIS', file, write_console)
self.rwd_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('STATE FREQUENCY ANALYSIS', file, write_console)
self.state_freq_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('STATE-ACTION FREQUENCY ANALYSIS', file, write_console)
self.state_action_freq_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('VALUE ANALYSIS', file, write_console)
self.value_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('RECENCY ANALYSIS', file, write_console)
self.recency_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('TRANSITION VALUE ANALYSIS', file, write_console)
self.trans_value_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('SEQUENCE ANALYSIS', file, write_console)
self.sequence_analysis._save_report(file, write_console)
print_line('\n===================================================================', file, write_console)
print_line('CONTRADICTION ANALYSIS', file, write_console)
self.contradiction_analysis._save_report(file, write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
action_names = self.config.get_action_names()
rwd_avg, rwd_std, n = self.avg_reward
print_line('====================================', file, write_console)
print_line(
'Average overall reward: {:.3f} ± {:.3f} (count: {})'.format(
rwd_avg, rwd_std, n), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} state-action outliers found (outlier threshold: {}):'.format(
len(self.state_action_outliers), self.state_outlier_stds),
file, write_console)
for s, a, rwd, n in self.state_action_outliers:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec)
print_line(
'\t{}-{} - {} (mean reward: {:.3f}, count: {})'.format(
s, feats_labels, action_names[a], rwd, n), file,
write_console)
print_line('====================================', file, write_console)
print_line('Actions\' average rewards:', file, write_console)
for a in range(len(self.action_rwds_avg)):
rwd_avg, rwd_std, n = self.action_rwds_avg[a]
print_line(
'\t{}: {:.3f} ± {:.3f} (count: {})'.format(
action_names[a], rwd_avg, rwd_std, n), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} feature-action outliers found (outlier threshold: {}):'.format(
len(self.feature_action_outliers), self.feature_outlier_stds),
file, write_console)
for f, v, a, rwd_avg, rwd_std, n in self.feature_action_outliers:
obs_vec = np.zeros(len(feats_nbins), np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line(
'\t{} - {} (avg. reward: {:.3f} ± {:.3f}, count: {})'.format(
feat_label, action_names[a], rwd_avg, rwd_std, n), file,
write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
action_names = self.config.get_action_names()
print_line(
'Min. state-action support: {}'.format(self.min_state_count), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} certain state transitions found (max. dispersion: {}):'.format(
len(self.certain_trans), self.certain_trans_max_disp), file,
write_console)
for s, a, supp, disp in self.certain_trans:
feats_labels = self.helper.get_features_labels(
get_features_from_index(s, feats_nbins), True)
print_line(
'\t({}-{}, {}) (supp: {}, disp: {:.3f})'.format(
s, feats_labels, action_names[a], supp, disp), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} uncertain state transitions found (min. dispersion: {}):'.
format(len(self.uncertain_trans),
self.uncertain_trans_min_disp), file, write_console)
for s, a, supp, disp in self.uncertain_trans:
feats_labels = self.helper.get_features_labels(
get_features_from_index(s, feats_nbins), True)
print_line(
'\t({}-{}, {}) (supp: {}, disp: {:.3f})'.format(
s, feats_labels, action_names[a], supp, disp), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} certain actions found (max. dispersion: {}):'.format(
len(self.certain_actions), self.certain_action_max_disp), file,
write_console)
for a, disp in self.certain_actions:
print_line('\t{} (disp: {:.3f})'.format(action_names[a], disp),
file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} uncertain actions found (min. dispersion: {}):'.format(
len(self.uncertain_actions), self.uncertain_action_min_disp),
file, write_console)
for a, disp in self.uncertain_actions:
print_line('\t{} (disp: {:.3f})'.format(action_names[a], disp),
file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} certain state feature-action pairs found (max disp: {}):'.
format(len(self.certain_feats),
self.certain_feat_max_disp), file, write_console)
num_feats = len(feats_nbins)
for f, v, a, disp in self.certain_feats:
obs_vec = np.zeros(num_feats, np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line(
'\t{} - {} (disp: {:.3f})'.format(feat_label, action_names[a],
disp), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} uncertain state feature-action pairs found (min. disp: {}):'.
format(len(self.uncertain_feats),
self.uncertain_feat_min_disp), file, write_console)
for f, v, a, disp in self.uncertain_feats:
obs_vec = np.zeros(num_feats, np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line(
'\t{} - {} (disp: {:.3f})'.format(feat_label, action_names[a],
disp), file, write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
action_names = self.config.get_action_names()
print_line('====================================', file, write_console)
print_line('{} total time-steps'.format(self.total_time_steps), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} earlier states found (min. support: {}, max time-step: {}):'.
format(len(self.earlier_states), self.min_state_count,
int(self.state_max_time_step * self.total_time_steps)),
file, write_console)
for s, t, n in self.earlier_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line(
'\t{}-{} (count: {}, last time-step: {})'.format(
s, feats_labels, n, t), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} earlier state-action pairs found (min. support: {}, max time-step: {}):'
.format(len(self.earlier_actions), self.min_state_count,
int(self.action_max_time_step * self.total_time_steps)),
file, write_console)
for s, a, t, n in self.earlier_actions:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line(
'\t{}-{} - {} (count: {}, last time-step: {})'.format(
s, feats_labels, action_names[a], n, t), file,
write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
action_names = self.config.get_action_names()
print_line('====================================', file, write_console)
print_line(
'Action JS divergence threshold: {} per action:'.format(
self.action_dist_div_threshold / self.config.num_actions),
file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} contradictory values states found):'.format(
len(self.contradictory_value_states)), file, write_console)
for s, n, jsd, diff_actions in self.contradictory_value_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
action_labels = [action_names[a] for a in diff_actions]
print_line(
'\t{}-{} (jsd: {:.3f}, count: {})'.format(
s, feats_labels, jsd, n), file, write_console)
print_line('\t\tDivergent actions: {}'.format(action_labels), file,
write_console)
self._print_actions('Rwd. dist', diff_actions,
self.state_rewards[s], file, write_console)
self._print_actions('Val. dist', diff_actions,
self.state_values[s], file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} contradictory count states found):'.format(
len(self.contradictory_count_states)), file, write_console)
for s, n, jsd, diff_actions in self.contradictory_count_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
action_labels = [action_names[a] for a in diff_actions]
print_line(
'\t{}-{} (jsd: {:.3f}, count: {})'.format(
s, feats_labels, jsd, n), file, write_console)
print_line('\t\tDivergent actions: {}'.format(action_labels), file,
write_console)
self._print_actions('Val. dist', diff_actions,
self.state_values[s], file, write_console)
self._print_actions('Count dist', diff_actions,
self.state_counts[s], file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} contradictory goal states found):'.format(
len(self.contradictory_goal_states)), file, write_console)
for s, n in self.contradictory_goal_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} (count: {})'.format(s, feats_labels, n), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} contradictory feature-action associations found):'.format(
len(self.contradictory_feature_actions)), file, write_console)
for f, v, a in self.contradictory_feature_actions:
obs_vec = np.zeros(len(feats_nbins), np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line('\t{} - {}'.format(feat_label, action_names[a]), file,
write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
action_names = self.config.get_action_names()
print_line('====================================', file, write_console)
print_line(
'Coverage of the state-action space (for visited states): {:.2f}%'.
format(self.state_action_coverage * 100.), file, write_console)
mean, std, n = self.mean_action_dispersion
print_line(
'Mean dispersion of the execution of actions in visited states: {:.3f} ± {:.3f} (count: {})'
.format(mean, std, n), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} certain states found (max. dispersion: {}):'.format(
len(self.certain_states), self.certain_state_max_disp), file,
write_console)
for s, disp, max_actions in self.certain_states:
max_action_labels = [action_names[a] for a in max_actions]
feats_labels = self.helper.get_features_labels(
get_features_from_index(s, feats_nbins), True)
print_line(
'\t{}-{} (mean disp.: {:.3f}, max actions {})'.format(
s, feats_labels, disp, max_action_labels), file,
write_console)
print_line('====================================', file, write_console)
print_line(
'{} uncertain states found (min. dispersion: {}):'.format(
len(self.uncertain_states), self.uncertain_state_min_disp),
file, write_console)
for s, disp in self.uncertain_states:
feats_labels = self.helper.get_features_labels(
get_features_from_index(s, feats_nbins), True)
print_line(
'\t{}-{} (mean disp.: {:.3f})'.format(s, feats_labels, disp),
file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} certain state features found (max. dispersion: {}):'.format(
len(self.certain_feats), self.certain_feat_max_disp), file,
write_console)
num_feats = len(feats_nbins)
for f, v, disp, max_actions in self.certain_feats:
max_action_labels = [action_names[a] for a in max_actions]
obs_vec = np.zeros(num_feats, np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line(
'\t{} (mean disp.: {:.3f}, max actions: {})'.format(
feat_label, disp, max_action_labels), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} uncertain state features found (min. dispersion: {}):'.format(
len(self.uncertain_feats), self.uncertain_feat_min_disp), file,
write_console)
for f, v, disp in self.uncertain_feats:
obs_vec = np.zeros(num_feats, np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line('\t{} (mean disp.: {:.3f})'.format(feat_label, disp),
file, write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
action_names = self.config.get_action_names()
mean, std, n = self.avg_value
print_line('====================================', file, write_console)
print_line(
'Average overall value: {:.3f} ± {:.3f} (count: {})'.format(
mean, std, n), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} state-action value outliers found (min. support: {}, outlier threshold: {}):'
.format(len(self.mean_val_state_action_outliers),
self.min_state_count, self.state_value_outlier_stds), file,
write_console)
for s, actions, val, n in self.mean_val_state_action_outliers:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
act_labels = [action_names[a] for a in actions]
print_line(
'\t{}-{} (value: {:.3f}, count: {})\n\t\tMax actions: {}'.
format(s, feats_labels, val, n,
act_labels), file, write_console)
mean, std, n = self.avg_pred_error
print_line('====================================', file, write_console)
print_line(
'Average overall prediction error: {:.3f} ± {:.3f} (count: {})'.
format(mean, std, n), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} state-action prediction error outliers found (min. support: {}, outlier threshold: {}):'
.format(len(self.pred_error_state_action_outliers),
self.min_state_count, self.pred_error_outlier_stds), file,
write_console)
for s, a, pred_error, n in self.pred_error_state_action_outliers:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line(
'\t{}-{} - {} (avg. prediction error: {:.3f}, count: {})'.
format(s, feats_labels, action_names[a], pred_error,
n), file, write_console)
print_line('====================================', file, write_console)
print_line('Actions\' average values:', file, write_console)
for a, avg, std in self.action_vals_avg:
print_line(
'\t{}: {:.3f} ± {:.3f}'.format(action_names[a], avg, std),
file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} feature value outliers found (outlier threshold: {}):'.format(
len(self.mean_val_feature_outliers),
self.feature_outlier_stds), file, write_console)
for f, v, avg, std in self.mean_val_feature_outliers:
obs_vec = np.zeros(len(feats_nbins), np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line(
'\t{} (avg. value: {:.3f} ± {:.3f})'.format(
feat_label, avg, std), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} feature pred. error outliers found (outlier threshold: {}):'.
format(len(self.pred_error_feature_outliers),
self.feature_outlier_stds), file, write_console)
for f, v, a, avg, std in self.pred_error_feature_outliers:
obs_vec = np.zeros(len(feats_nbins), np.uint32)
obs_vec[f] = v
feat_label = self.helper.get_features_labels(obs_vec)[f]
print_line(
'\t{} - {} (avg. pred. error: {:.3f} ± {:.3f})'.format(
feat_label, action_names[a], avg, std), file,
write_console)
def _save_report(self, file, write_console):
feats_nbins = self.helper.get_features_bins()
print_line('====================================', file, write_console)
print_line('{} local minima states found (min. transition support: {}):'.format(
len(self.local_minima_states), self.min_transition_count), file, write_console)
for s, val, ns_val_avg, n in self.local_minima_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} (count: {}, value: {:.3f} <= avg next values: {:.3f})'.format(
s, feats_labels, n, val, ns_val_avg), file, write_console)
print_line('====================================', file, write_console)
print_line('{} absolute minima states found (min. transition support: {}):'.format(
len(self.absolute_minima_states), self.min_transition_count), file, write_console)
for s, val, ns_val_avg, n in self.absolute_minima_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} (count: {}, value: {:.3f} <= avg next values: {:.3f})'.format(
s, feats_labels, n, val, ns_val_avg), file, write_console)
print_line('====================================', file, write_console)
print_line('{} local maxima states found (min. transition support: {}):'.format(
len(self.local_maxima_states), self.min_transition_count), file, write_console)
for s, val, ns_val_avg, n in self.local_maxima_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} (count: {}, value: {:.3f} >= avg next values: {:.3f})'.format(
s, feats_labels, n, val, ns_val_avg), file, write_console)
print_line('====================================', file, write_console)
print_line('{} absolute maxima states found (min. transition support: {}):'.format(
len(self.absolute_maxima_states), self.min_transition_count), file, write_console)
for s, val, ns_val_avg, n in self.absolute_maxima_states:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} (count: {}, value: {:.3f} >= avg next values: {:.3f})'.format(
s, feats_labels, n, val, ns_val_avg), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} value difference mean state outliers found (min. transition support: {}, outlier stds: {}):'.format(
len(self.val_diff_mean_action_outliers), self.min_transition_count, self.state_diff_outlier_stds),
file, write_console)
action_names = self.config.get_action_names()
for s, a, diff, n in self.val_diff_mean_action_outliers:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} - {} (trans. count: {}, mean value diff.: {:.3f})'.format(
s, feats_labels, action_names[a], n, diff), file, write_console)
print_line('====================================', file, write_console)
print_line(
'{} value difference variance state-action outliers found (min. transition support: {}, outlier stds: {}):'.
format(len(self.val_diff_variance_state_outliers), self.min_transition_count,
self.state_diff_outlier_stds), file, write_console)
for s, diff, n in self.val_diff_variance_state_outliers:
obs_vec = get_features_from_index(s, feats_nbins)
feats_labels = self.helper.get_features_labels(obs_vec, True)
print_line('\t{}-{} (count: {}, value diff. variance: {:.3f})'.format(
s, feats_labels, n, diff), file, write_console)
# creates log file
log_file = open(join(output_dir, 'log.txt'), 'w')
print()
# ============================================================
# PART 1 : collects performance stats across trials/episodes per agent
subdir = join(output_dir, 'performance')
clear_dir(subdir)
stats = OrderedDict()
var_names = list(helpers[0][0].stats_collector.all_variables())
for ag_idx, ag_helpers in enumerate(helpers):
ag_name = ag_names[ag_idx]
print_line(
'Collecting stats across {} trials of {} episodes for agent \'{}\'...'
.format(len(ag_helpers), ag_helpers[0].config.num_episodes,
ag_name), log_file)
# saves / copies configs to file
ag_helpers[0].config.save_json(
join(output_dir, 'config-{}.json'.format(ag_name.lower())))
# collects stats for all trials
stats[ag_name] = {}
stats_collectors = [helper.stats_collector for helper in ag_helpers]
for var_name in var_names:
ag_stats = StatsCollector.get_mean_trials_stats(
stats_collectors, var_name, ag_helpers[0].config.num_episodes)
stats[ag_name][var_name] = ag_stats
# prints comparison (csv, screen, bar-chart) between agents for the several variables
