Iterative Optimization Heuristics Profiler

The performance analyzer for Iterative Optimization Heuristics (IOHs). It provides:

• a web-based interface to analyze and visualize the empirical performance of IOHs
• interactive plot
• statistical evaluation
• report generation
• R programming interfaces in the backend

Development Team

• Hao Wang, Leiden Institute of Advanced Computer Science.
• Diederick Vermetten, Leiden Institute of Advanced Computer Science.
• Furong Ye, Leiden Institute of Advanced Computer Science.
• Carola Doerr, CNRS and Sorbonne University.
• Ofer Shir Migal - The Galilee Research Institute, Tel-Hai College.
• Thomas Bäck, Leiden Institute of Advanced Computer Science.

Installation

Software dependency

• [mandatory] R As IOHanalyzer is written as a R package, the R environment has to be installed first. The binary file and installation manual for R can be found here https://cran.r-project.org/.
• [optional] orca required to download plotly figures. Please see https://github.com/plotly/orca for the installation instruction.
• [optional] inkscape required to support pdf and eps figure format in downloading. Please visit the Inskscape Wiki page http://wiki.inkscape.org/wiki/index.php/Installing_Inkscape for the detail.

Host it online yourself?

We provide docker file for deploying IOHanalyzer on the server. Please see https://github.com/IOHprofiler/IOHanalyzer-docker for details.

Reference

When using IOHprofiler and parts thereof, please kindly cite this work as

Carola Doerr, Hao Wang, Furong Ye, Sander van Rijn, Thomas Bäck: IOHprofiler: A Benchmarking and Profiling Tool for Iterative Optimization Heuristics, arXiv e-prints:1810.05281, 2018.

@ARTICLE{IOHprofiler,
author = {Carola Doerr and Hao Wang and Furong Ye and Sander van Rijn and Thomas B{\"a}ck},
title = {{IOHprofiler: A Benchmarking and Profiling Tool for Iterative Optimization Heuristics}},
journal = {arXiv e-prints:1810.05281},
archivePrefix = "arXiv",
eprint = {1810.05281},
year = 2018,
month = oct,
keywords = {Computer Science - Neural and Evolutionary Computing},
url = {https://arxiv.org/abs/1810.05281}
}


Acknowledgements

This work was supported by the Chinese scholarship council (CSC No. 201706310143), a public grant as part of the Investissement d’avenir project, reference ANR-11-LABX-0056-LMH, LabEx LMH, in a joint call with the Gaspard Monge Program for optimization, operations research, and their interactions with data sciences, by Paris Ile-de-France Region, and by COST Action CA15140 “Improving Applicability of Nature-Inspired Optimisation by Joining Theory and Practice (ImAppNIO)”.

This application is governed by the BSD 3-Clause license.

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

• Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.

• Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.

• Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

Welcome to IOHanalyzer!

This the analyzer for the empirical performance of Iterative Optimization Heuristics (IOHs). It provides two perspectives to look at the empirical performance:

• Fixed-Target running time: focuses on the distribution of running time (a.k.a. the number of function evaluations) at various given target values (thus called “fixed-target”), functions and dimensions.
• Fixed-Budget results: focuses on the distribution of the best function value at various given budget values (thus called “fixed-budget”), functions and dimensions.

To get started, you could

• upload your own dataset using the Upload Data box on the bottom left. The supported format is specified in the data format tab.
• or choose one of the pre-loaded datasets in the Load Data from Repository box on the bottom right and directly explore the analyzer.

In details, the following functionalities are provided:

Data Summary Expected Values Probability Density/Mass Function Empirical Cumulative Distribution (ECDF)
Fixed-Target running time Summary of running times Expected Running Time (ERT) Probability Mass Function of running time ECDF of running times
Fixed-Budget results Summary of function values Expected Target Value Probability Density Function of target values ECDF of targets times

For more information on the features of IOHanalyzer and how to use them, as well as a full specification of the accepted data formats, please visit our wiki.

IOHexperimenter, Nevergrad, and BBOB/COCO data files are accepted. For alternative data files, please convert them to the format described here.

When the dataset is huge, the alignment can take a very long time. In this case, you could toggle the efficient mode to subsample the dataset. However, the precision of data will be compromised.

Load the data from the available repositories. There are currently three available sources:

• Data generated with the PBO-suite, implemented in the IOHexperimenter
• All data generated by the nevergrad benchmarking framework
• The majority of the publicly available benchmark data on the single-objective BBOB framework

Data Processing Prompt

Overview of Selected Function

Select which algorithms to show.

Data Overview

Select which algorithms to show.

This table provides an overview on function values for all algorithms chosen on the left:

• worst recorded: the worst $$f(x)$$ value ever recorded across all iterations,
• worst reached: the worst $$f(x)$$ value reached in the last iterations,
• best reached: the best $$f(x)$$ value reached in the last iterations,
• mean reached: the mean $$f(x)$$ value reached in the last iterations,
• median reached: the median $$f(x)$$ value reached in the last iterations,
• succ: the number of runs which successfully hit the best reached $$f(x)$$.

Runtime Statistics at Chosen Target Values

Set the range and the granularity of the results. The table will show fixed-target runtimes for evenly spaced target values.

This table summarizes for each algorithm and each target value chosen on the left:

• runs: the number of runs that have found at least one solution of the required target quality $$f(x)$$,
• mean: the average number of function evaluations needed to find a solution of function value at least $$f(x)$$
• median, $$2\%, 5\%,\ldots,98\%$$ : the quantiles of these first-hitting times

When not all runs managed to find the target value, the statistics hold only for those runs that did. That is, the mean value is the mean of the successful runs. Same for the quantiles. An alternative version with simulated restarts is currently in preparation.

Original Runtime Samples

Set the range and the granularity of the results. The table will show fixed-target runtimes for evenly spaced target values.

This table shows for each selected algorithm $$A$$, each selected target value $$f(x)$$, and each run $$r$$ the number $$T(A,f(x),r)$$ of evaluations performed by the algorithm until it evaluated for the first time a solution of quality at least $$f(x)$$.

Expected Runtime (ERT): single function

Range of the displayed target values

The mean, median, standard deviation and ERT of the runtime samples are depicted against the best objective values. The displayed elements (mean, median, standard deviations and ERT) can be switched on and off by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Expected Runtime Comparisons (across dimensions)

The ERT of the runtime samples across all functions. ERT is decided based on the target values in the table below, with the default being the best reached f(x) by any of the selected algorithms. Infinite ERTS are shown as seperate dots on the graph.

The chosen target values per dimension are as follows (double click an entry to edit it):

The raw ERT-values are:

Expected Runtime Comparisons (across functions on one dimension)

The ERT of the runtime samples across all functions. ERT is decided based on the target values in the table below, with the default being the best reached f(x) by any of the selected algorithms. Infinite ERTS are shown as seperate dots on the graph.

The chosen target values per function are as follows (double click an entry to edit it):

The raw ERT-values are:

Expected Runtime (ERT): all functions

The ERT is shown against the target values for all functions in the selected dimension.

Histogram of Fixed-Target Runtimes

Choose whether the histograms are overlaid in one plot or separated in several subplots:

This histogram counts how many runs needed between $$t$$ and $$t+1$$ function evaluations. The bins $$[t,t+1)$$ are chosen automatically. The bin size is determined by the so-called Freedman–Diaconis rule: $$\text{Bin size}= 2\frac{Q_3 - Q_1}{\sqrt[3]{n}}$$, where $$Q_1, Q_3$$ are the $$25\%$$ and $$75\%$$ percentile of the runtime and $$n$$ is the sample size. The displayed algorithms can be selected by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Empirical Probability Mass Function of the Runtime

Warning! The probability mass function of the runtime is approximated by the treating the runtime as a continuous random variable and applying the kernel estimation (KDE):

The plot shows the distribution of the first hitting times of the individual runs (dots), and an estimated distribution of the probability mass function. The displayed algorithms can be selected by clicking on the legend on the right. A tooltip and toolbar appear when hovering over the figure. This also includes the option to download the plot as png file. A csv file with the runtime data can be downlowaded from the Data Summary tab.

Empirical Cumulative Distribution: Single target

Select the target values for which EDCF curves are displayed

Each EDCF curve shows the proportion of the runs that have found a solution of at least the required target value within the budget given by the $$x$$-axis. The displayed curves can be selected by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure. This also includes the option to download the plot as png file.

Aggregated Empirical Cumulative Distribution: Single function

Set the range and the granularity of the quality targets taken into account in the ECDF curve. The plot will show the ECDF curves for evenly spaced target values.

The evenly spaced target values are:

The fraction of (run,target value) pairs $$(i,v)$$ satisfying that the best solution that the algorithm has found in the $$i$$-th run within the given time budget $$t$$ has quality at least $$v$$ is plotted against the available budget $$t$$. The displayed elements can be switched on and off by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Area Under the ECDF

Set the range and the granularity of the evenly spaced quality targets taken into account in the plot.

The area under the ECDF is caculated for the sequence of target values specified on the left. The displayed values are normalized against the maximal number of function evaluations for each algorithm. Intuitively, the larger the area, the better the algorithm. The displayed algorithms can be selected by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure. This also includes the option to download the plot as png file.

Aggregated Empirical Cumulative Distribution: All functions

Alternatively, you can download the table containing the target values for each (function, dimension)-pair and edit the table as you want. Please keep the file format when modifying it.

The fraction of (run,target value, ...) pairs $$(i,v, ...)$$ satisfying that the best solution that the algorithm has found in the $$i$$-th (run of function $$f$$ in dimension $$d$$) within the given time budget $$t$$ has quality at least $$v$$ is plotted against the available budget $$t$$. The displayed elements can be switched on and off by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure. Aggregation over functions and dimension can be switched on or off using the checkboxes on the left; when aggregation is off the selected function / dimension is chosen according the the value in the bottom-left selection-box.

The selected targets are:

Parameter Statistics at Chosen Target Values

Set the range and the granularity of the results. The table will show fixed-target parameter values for evenly spaced target values.

This table summarizes for each algorithm and each target value chosen on the left:

• runs: the number of runs where non-missing parameter values are observed, for each required target value $$f(x)$$,
• mean: the average value of the specified parameter when target value $$f(x)$$ is hit.
• median, $$2\%, 5\%,\ldots,98\%$$ : the quantiles of these parameter values

When not all runs managed to find the target value, the statistics hold only for those runs that did. That is, the mean value is the mean of the successful runs. Same for the quantiles. An alternative version with simulated restarts is currently in preparation.

Expected Parameter Value (per function)

Range of the function values ($$x$$ axis)

The mean or median of internal parameters of the algorithm found with a fixed-budget of evaluations are depicted against the budget. The displayed elements can be switched on and off by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Parameter Sample at Chosen Target Values

Set the range and the granularity of the results. The table will show fixed-target parameter values for evenly spaced target values.

This table shows for each selected algorithm $$A$$, each selected target value $$f(x)$$, and each run $$r$$ the parameter value observed when the target value $$f(x)$$ is reached for the first time.

Hypothesis Testing

The Kolmogorov-Smirnov test is performed on empirical CDFs of running times for each pair of algorithms, in order to determine which algorithm gives a significantly smaller running time distribution. The resulting p-values are arranged in a matrix, where each cell $$(i, j)$$ contains a p-value from the test with the alternative hypothesis: the running time of algorithm $$i$$ is smaller (thus better) than that of $$j$$.

Decisions of the test, based on the $$p-$$ value matrix and the $$\alpha$$ value, are visualized in a heatmap (left) and a network (right). In each cell (row, column) of the heatmap, the alternative hypothesis is again: the row algorithm has smaller (better) running time than the column. The color indicates:

• Red: A row is better than the column
• Blue: A row is worse than the column
• Gray: no significant distinction between the row and column
On the right subplot, the partial order resulted from the test is visualized as a network, where an arrow from algorithm $$A$$ to $$B$$ indicates $$A$$ is siginicantly better than $$B$$ with the specified $$\alpha$$ value.

Glicko2-based ranking

The Glicko2 This procedure ranks algorithms based on a glico2-procedure. Every round, for every function and dimension of the datasetlist, each pair of algorithms competes. This competition samples a random runtime for the provided target (best achieved target among all algorithms). Whichever algorithm has the lower runtime wins the game. Then, from these games, the glico2-rating is used to determine the ranking.

The chosen target function values per (function, dimension)-pair are as follows (double click an entry to edit it):

The results of the ranking are:

Data Overview

Select which algorithms to show.

Target Statistics at Chosen Budget Values

Set the range and the granularity of the results. The table will show function values that have been reached within evenly spaced evaluation budgets.

This table summarizes for each algorithm and each budget $$B$$ chosen on the left:

• runs: the number of runs that have performed at least $$B$$ evaluations,
• mean: the average best-so-far function value obtain within a budget of $$B$$ evaluations,
• median, $$2\%, 5\%,\ldots,98\%$$ : the quantiles of the best function values found within the first $$B$$ evaluations.

When not all runs evaluated at least $$B$$ search points, the statistics hold for the subset of runs that did. Alternative statistics using simulated restarted algorithms are in preparation.

Original Target Samples

Set the range and the granularity of the results. The table will show function values that have been reached within evenly spaced evaluation budgets.

This table shows for each selected algorithm $$A$$, each selected target value $$f(x)$$, and each run $$r$$ the number $$T(A,f(x),r)$$ of evaluations performed by the algorithm until it evaluated for the first time a solution of quality at least $$f(x)$$.

Expected Target Value (per function)

Range of the displayed budget values

The mean, median and standard deviation of the best function values found with a fixed-budget of evaluations are depicted against the budget. The displayed elements can be switched on and off by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Expected Function Value comparisons

The chosen budget values per function are as follows (double click an entry to edit it):

The raw function-values are:

Histogram of Fixed-Budget Targets

Choose whether the histograms are overlaid in one plot or separated in several subplots:

This histogram counts the number of runs whose best-so-far function values within the first $$B$$ evaluations is between $$v_i$$ and $$v_{i+1}$$. The buckets $$[v_i,v_{i+1})$$ are chosen automatically according to the so-called Freedman–Diaconis rule: $$\text{Bin size}= 2\frac{Q_3 - Q_1}{\sqrt[3]{n}}$$, where $$Q_1, Q_3$$ are the $$25\%$$ and $$75\%$$ percentile of the runtime and $$n$$ is the sample size. The displayed algorithms can be selected by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Empirical Probability Density Function of Fixed-Budget Function Values

Select the budget for which the distribution of best-so-far function values is shown

The plot shows, for the budget selected on the left, the distribution of the best-so-far function values of the individual runs (dots), and an estimated distribution of the probability mass function. The displayed algorithms can be selected by clicking on the legend on the right. A tooltip and toolbar appear when hovering over the figure. A csv file with the runtime data can be downloaded from the Data Summary tab.

Empirical Cumulative Distribution of the Fixed-Budget Values: Single Budgets

Select the budgets for which EDCF curves are displayed

Each EDCF curve shows the proportion of the runs that have found within the given budget B a solution of at least the required target value given by the x-axis. The displayed curves can be selected by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Empirical Cumulative Distribution of the Fixed-Budget Values: Aggregation

Set the range and the granularity of the budgets taken into account in the ECDF curve. The plot will show the ECDF curves for evenly spaced budgets.

The evenly spaced budget values are:

The fraction of (run,budget) pairs $$(i,B)$$ satisfying that the best solution that the algorithm has found in the $$i$$-th run within the first $$B$$ evaluations has quality at most $$v$$ is plotted against the target value $$v$$. The displayed elements can be switched on and off by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Area Under the ECDF

Set the range and the granularity of the evenly spaced budgets.

The area under the ECDF is caculated for the sequence of budget values specified on the left. The displayed values are normalized against the maximal target value recorded for each algorithm. Intuitively, the smaller the area, the better the algorithm. The displayed algorithms can be selected by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Expected Parameter Value (per function)

Range of the budget values ($$x$$ axis)

The mean or median of internal parameters of the algorithm found with a fixed-budget of evaluations are depicted against the budget. The displayed elements can be switched on and off by clicking on the legend on the right. A tooltip and toolbar appears when hovering over the figure.

Parameter Statistics at Chosen Runtime Values

Set the range and the granularity of the results. The table will show fixed-budget parameter values for evenly spaced budget values.

This table summarizes for each algorithm and each runtime value chosen on the left:

• runs: the number of runs where non-missing parameter values are observed, for each required runtime value,
• mean: the average value of the specified parameter after the specified amount of function evaluations.
• median, $$2\%, 5\%,\ldots,98\%$$ : the quantiles of these parameter values

When not all runs used the specified runtime value, the statistics hold only for those runs that did. That is, the mean value is the mean of the applicable runs. Same for the quantiles.

Parameter Sample at Chosen Runtime Values

Set the range and the granularity of the results. The table will show fixed-budget parameter values for evenly spaced budget values.

This table shows for each selected algorithm $$A$$, each selected runtime value $$b$$, and each run $$r$$ the parameter value observed when the after a total budget of $$b$$ function evaluations has been used.

Hypothesis Testing

The Kolmogorov-Smirnov test is performed on empirical CDFs of running times for each pair of algorithms, in order to determine which algorithm gives a significantly smaller running time distribution. The resulting p-values are arranged in a matrix, where each cell $$(i, j)$$ contains a p-value from the test with the alternative hypothesis: the running time of algorithm $$i$$ is smaller (thus better) than that of $$j$$.

Decisions of the test, based on the $$p-$$ value matrix and the $$\alpha$$ value, are visualized in a heatmap (left) and a network (right). In each cell (row, column) of the heatmap, the alternative hypothesis is again: the row algorithm has smaller (better) running time than the column. The color indicates:

• Red: A row is better than the column
• Blue: A row is worse than the column
• Gray: no significant distinction between the row and column
On the right subplot, the partial order resulted from the test is visualized as a network, where an arrow from algorithm $$A$$ to $$B$$ indicates $$A$$ is siginicantly better than $$B$$ with the specified $$\alpha$$ value.

Glicko2-based ranking

The Glicko2 This procedure ranks algorithms based on a glico2-procedure. Every round, for every function and dimension of the datasetlist, each pair of algorithms competes. This competition samples a random function value for the provided runtime (maximum used among all algorithms). Whichever algorithm has the better function value wins the game. Then, from these games, the glico2-rating is used to determine the ranking.

The chosen budget values per (function, dimension)-pair are as follows (double click an entry to edit it):

The results of the ranking are:

Parallel coordinate plot of final position

The location of the best found solution by the algorithm, for each of the runs. Each x-value corresponds to one coordinate of the solution.

Color Settings

Example of the current colorscheme.

General settings

Set the general properties