[cocopp] two budget crosses in runtime distributions

Author: nikohansen

code-postprocessing/cocopp/compall/pprldmany.py

@@ -52,7 +52,9 @@
 save_zoom = False  # save zoom into left and right part of the figures
 perfprofsamplesize = genericsettings.simulated_runlength_bootstrap_sample_size  # number of bootstrap samples drawn for each fct+target in the performance profile
 nbperdecade = 1
-median_max_evals_marker_format = ['x', 12.5, 1]  # [symbol, size, edgewidth]
+max_evals_marker_format = ['x', 12.5, 1]  # [symbol, size, edgewidth]
+max_evals_single_marker_format = ['+', 14, 1]  # [symbol, size, edgewidth]
+max_evals_percentile = 90
 label_fontsize = 15  # was 17
 xticks_fontsize = 16
 yticks_fontsize = 14
@@ -195,8 +197,7 @@ def beautify():
     plt.ylim(-0.0, 1.0)
 
 
-
-def plotdata(data, maxval=None, maxevals=None, CrE=0., **kwargs):
+def plotdata(data, maxval=None, maxevals=None, CrE=0., maxevals2=None, **kwargs):
     """Draw a normalized ECDF. What means normalized?
     
     :param seq data: data set, a 1-D ndarray of runlengths
@@ -206,7 +207,9 @@ def plotdata(data, maxval=None, maxevals=None, CrE=0., **kwargs):
     :param seq maxevals: if provided, will plot the median of this
                          sequence as a single cross marker
     :param float CrE: Crafting effort the data will be multiplied by
-                      the exponential of this value.
+                      the exponential of this value
+    :param maxevals2: a single value or values to be plotted as median(maxevals2)
+                      with the same marker as maxevals
     :param kwargs: optional arguments provided to plot function.
     
     """
@@ -250,33 +253,37 @@ def plotdata(data, maxval=None, maxevals=None, CrE=0., **kwargs):
                                         logscale=False, clip_on=False, **kwargs)
         # res = plotUnifLogXMarkers(x2, y2, nbperdecade, logscale=False, **kwargs)
 
-    if maxevals:  # Should cover the case where maxevals is None or empty
-        x3 = np.median(maxevals)  # change it only here
-        if (x3 <= maxval and
-            # np.any(x2 <= x3) and   # maxval < median(maxevals)
-            not plt.getp(res[-1], 'label').startswith('best')
-            ): # TODO: HACK for not considering a "best" algorithm line
-            # Setting y3
-            if n == 0:
-                y3 = 0
-            else:
-                try:
-                    y3 = y2[x2 <= x3][-1]  # find right y-value for x3==median(maxevals)
-                except IndexError:  # median(maxevals) is smaller than any data, can only happen because of CrE?
-                    y3 = y2[0]
-            h = plt.plot((x3,), (y3,),
-                         marker=median_max_evals_marker_format[0],
-                         markersize=median_max_evals_marker_format[1] * size_correction_from_n_foreground**0.85,
-                         markeredgewidth=median_max_evals_marker_format[2],
-                         # marker='x', markersize=24, markeredgewidth=3, 
-                         markeredgecolor=plt.getp(res[0], 'color'),
-                         ls=plt.getp(res[0], 'linestyle'),
-                         color=plt.getp(res[0], 'color'),
-                         # zorder=1.6   # zorder=0;1;1.5 is behind the grid lines, 2 covers other lines, 1.6 is between
-                         )
-            h.extend(res)
-            res = h  # so the last element in res still has the label.
-            # Only take sequences for x and y!
+        for maxeval_, format in ((maxevals, max_evals_marker_format),
+                                 (maxevals2, max_evals_single_marker_format)):
+            if not maxeval_:  # cover the case where maxevals is None or empty
+                continue
+            x3 = np.median(maxeval_)  # change it only here
+            if (x3 <= maxval and
+                # np.any(x2 <= x3) and   # maxval < median(maxevals)
+                not plt.getp(res[-1], 'label').startswith('best')
+                ): # TODO: HACK for not considering a "best" algorithm line
+                # Setting y3
+                if n == 0:
+                    y3 = 0
+                else:
+                    try:
+                        y3 = y2[x2 <= x3][-1]  # find right y-value for x3==median(maxevals)
+                    except IndexError:  # median(maxevals) is smaller than any data, can only happen because of CrE?
+                        y3 = y2[0]
+                h = plt.plot((x3,), (y3,),
+                            marker=format[0],
+                            markersize=format[1] * size_correction_from_n_foreground**0.85,
+                            markeredgewidth=format[2],
+                            # marker='x', markersize=24, markeredgewidth=3, 
+                            markeredgecolor=plt.getp(res[0], 'color'),
+                            ls=plt.getp(res[0], 'linestyle'),
+                            color=plt.getp(res[0], 'color'),
+                            # zorder=1.6   # zorder=0;1;1.5 is behind the grid lines, 2 covers other lines, 1.6 is between
+                            )
+                # h.extend(res)
+                # res = h  # so the last element in res still has the label.
+
+                # Only take sequences for x and y!
 
     return res
 
@@ -644,7 +651,8 @@ def main(dictAlg, order=None, outputdir='.', info='default',
             print('Crafting effort for', alg, 'is', CrE)
 
     dictData = {}  # list of (ert per function) per algorithm
-    dictMaxEvals = collections.defaultdict(list)  # list of (maxevals per function) per algorithm
+    dictMaxEvals = collections.defaultdict(list)  # sum(maxevals) / max(1, #success) per instance
+    dictMaxEvals2 = collections.defaultdict(list)  # max of successf and unsucc 90%tile runtime over all instances
 
     # funcsolved = [set()] * len(targets) # number of functions solved per target
     xbest = []
@@ -690,7 +698,8 @@ def main(dictAlg, order=None, outputdir='.', info='default',
             samplesize = int(samplesize)
         for f, dictAlgperFunc in sorted(dictFunc.items()):
             # print(target_values((f, dim)))
-            for j, t in enumerate(target_values((f, dim))):
+            targets = target_values((f, dim))
+            for j, t in enumerate(targets):
                 # for j, t in enumerate(testbedsettings.current_testbed.ecdf_target_values(1e2, f)):
                 # funcsolved[j].add(f)
 
@@ -737,11 +746,17 @@ def main(dictAlg, order=None, outputdir='.', info='default',
                         keyValue = 'f%d' % (f)
                     dictData.setdefault(keyValue, []).extend(x)
                     # dictMaxEvals.setdefault(keyValue, []).extend(runlengthunsucc)
-                    if len(runlengthunsucc):
-                        maxmed = np.median(runlengthunsucc)
-                        if len(runlengthsucc):
-                            maxmed = max((maxmed, np.median(runlengthsucc)))
-                        dictMaxEvals[keyValue].append(maxmed)
+                    if len(runlengthunsucc) and t == min(targets):  # only once, not for each target as it was before June 2024
+                        def percentile(vals, which=max_evals_percentile):
+                            return toolsstats.prctile(vals, [which])[0]
+                        if 1 < 3:  # budget_estimates gives the cross somewhat way too far to the right
+                            if 'entry' in locals():
+                                dictMaxEvals[keyValue].append(percentile(list(entry.budget_effective_estimates.values())))
+                        if 1 < 3:
+                            maxmed = percentile(runlengthunsucc)
+                            if len(runlengthsucc):
+                                maxmed = max((maxmed, percentile(runlengthsucc)))
+                            dictMaxEvals2[keyValue].append(maxmed)
 
             displaybest = plotType == PlotType.ALG
             if displaybest:
@@ -806,7 +821,6 @@ def algname_to_label(algname, dirname=None):
         for i, alg in enumerate(plotting_style.algorithm_list):
             try:
                 data = dictData[alg]
-                maxevals = dictMaxEvals[alg]
             except KeyError:
                 continue
 
@@ -837,7 +851,8 @@ def algname_to_label(algname, dirname=None):
 
             args.update(plotting_style.pprldmany_styles)  # no idea what this does, maybe update for background algorithms?
 
-            lines.append(plotdata(np.array(data), x_limit, maxevals,
+            lines.append(plotdata(np.array(data), x_limit,
+                                  dictMaxEvals[alg], maxevals2=dictMaxEvals2[alg],
                                   CrE=CrEperAlg[alg], **args))
 
     if 11 < 3:

code-postprocessing/cocopp/pproc.py

@@ -584,6 +584,7 @@ class DataSet(object):
         algId
         algs
         bootstrap_sample_size
+        budget_effective_estimates
         comment
         ...
         dim
@@ -598,6 +599,7 @@ class DataSet(object):
         ...
         info
         info_str
+        instance_index_lists
         instance_multipliers
         instancenumbers
         isBiobjective
@@ -619,8 +621,10 @@ class DataSet(object):
         reference_values
         splitByTrials
         success_ratio
+        successes_by_instance
         suite_name
         target
+        trial_count_by_instance
         >>> all(ds.evals[:, 0] == ds.target)  # first column of ds.evals is the "target" f-value
         True
         >>> # investigate row 0,10,20,... and of the result columns 0,5,6, index 0 is ftarget
@@ -2098,6 +2102,69 @@ def median_evals(self, target_values=None, append_instances=True):
         m[~np.isfinite(m)] = np.nan
         return m
 
+    def instance_index_lists(self, raw_values=True):
+        """return `OrderedDict` of index lists for each instance.
+
+        The index starts with 0 conforming with ``instancenumbers``,
+        ``maxevals``, `detEvals` and others. However in the ``evals``
+        array, column 0 contains f-values and the instance indices start
+        with 1.
+        """
+        if raw_values is not True:
+            raise NotImplementedError('instance index lists is not implemented for expanded evals')
+        res = collections.OrderedDict()
+        for index, i in enumerate(self.instancenumbers):
+            if i not in res:
+                res[i] = []
+            res[i] += [index]
+        return res
+
+    @property
+    def _budget_estimates(self):
+        """return `OrderedDict` of sum(maxevals) for each (raw data) instance.
+
+        This was implemented but never used.
+        """
+        res = collections.OrderedDict()
+        for instance, indices in self.instance_index_lists(raw_values=True).items():
+            res[instance] = sum(self.maxevals[i] for i in indices)
+            # res[instance] = max((max(s), sum(u)))
+            # res[instance] = sum(u) + (np.median(s) if s else 0)
+            # res[instance] = sum(u) + (max(s) if s else 0)
+        return res
+
+    @property
+    def budget_effective_estimates(self):
+        """return `OrderedDict` of ``sum(maxevals) / max(1, #successes)``
+
+        for each instance. This is similar to the budget of the
+        within-trial restarted algorithm and also equals to the
+        within-instance ERT for the most difficult target
+        ``self.precision`` when #successes > 0.
+        """
+        res = collections.OrderedDict()
+        successes = self.successes_by_instance()
+        for instance, indices in self.instance_index_lists(raw_values=True).items():
+            res[instance] = sum(self.maxevals[i] for i in indices
+                                ) / max((1, successes[instance]))
+        return res
+
+    def successes_by_instance(self, target=None, raw_values=True):
+        """return `OrderedDict` with number of successes for each instance"""
+        res = collections.OrderedDict()
+        try:
+            target = self.precision
+        except AttributeError:  # biobj case
+            target = 1e-8  # FIXME: is there a smarter way here?
+        evals = self.detEvals([target])[0]
+        for instance, indices in self.instance_index_lists(raw_values).items():
+            res[instance] = sum(np.isfinite(evals[i]) for i in indices)
+        return res
+
+    def trial_count_by_instance(self, target=None):
+        """return `Counter` `dict` with number of trials for each instance"""
+        return collections.Counter(self.instancenumbers)
+
     def _data_differ(self, ds):
         """return a list of targets for which `ds` differs from `self`