Skip to content

Latest commit

 

History

History
141 lines (122 loc) · 6.04 KB

README.md

File metadata and controls

141 lines (122 loc) · 6.04 KB
Raw

MonAPMC simulation experiments

This repository contains the numerical experiments to test the algorithm MonAPMC described in the companion paper [REF]. This algorithm is an attempt to implement an approximate Bayesian computation (ABC) algorithm that scales well with the number of computing cores available to perform simulations and the complexity of the simulation model. It is essentially the algorithm APMC [REF] equipped with a specific parallelization scheme.

The next section explains how to rerun the simulation experiments and reproduce the figures described in the rest of the document.

The ABC algorithms used in these experiments and the statistical computations are written in haskell, under the directory <./formulas>. The figures are generated with gnuplot and relevant gnuplot scripts are located in <./report>; the corresponding image files are in <./output/report>. One set of simulation that is particularly demanding computationally (see Run Comp LHS below) is run on the European Grid Infrastructure through the virtual organization vo.complex-systems.eu, using the software OpenMOLE. The openmole script is the file openmole/lhs.oms.

A consistent naming scheme accross the companion paper, the haskell code, gnuplot script files and the Makefile helps keeping track of where the various definitions in the text are defined in code and data files.

Directory structure

.
├── formulas: The haskell code implementing the algorithms, simulations and statistics.
│   ├── app: code of the haskell executables used to run the simulations and statistics.
│   ├── src
│   │   └── ABC: Implementations of MonAPMC and APMC algorithms
│   └── test: Tests
├── openmole: files to run the LHS experiment in openmole.
├── output
│   ├── formulas
│   │   ├── figure_data: input data for figures
│   │   │   ├── histo_run
│   │   │   ├── histo_steps
│   │   │   └── l2_vs_time
│   │   ├── repli: simulation results.
│   │   │   ├── comp
│   │   │   ├── run
│   │   │   └── steps
│   │   ├── run: simulation results.
│   │   ├── run_comp_lhs: simulation results.
│   │   └── steps: simulation results.
│   └── report: generated figures.
├── report: gnuplot scripts generating the figures.
└── sentinel: sentinel files (used by the Makefile).

Usage: how to reproduce the results?

Dependencies:

Both executables need to be on the path.

The repository uses make to run the simulations, compute statistics and generate the figures automatically.

To run the simulation experiment defined by the OpenMOLE script openmole/lhs.oms, you will need either to have credentials to use the European Grid Infrastructure (EGI) through the virtual organization vo.complex-systems.eu or use another execution environment by changing the env value in the script.

You may want to run openmole manually rather than through make to access openmole's web interface, through which you can watch the simulation's progression on the distant computing cluster. To do that, ensure that all the prerequisites are built with make openmole/formulas.container.tgz, then run the script <./openmole/lhs.oms> yourself using openmole.

Definitions

Here is a list of definitions given in the companion paper. This lists will help the reader make the correspondance between the paper and this repository by specifying where each definition is defined as code, stored as data or as an image file when appropriate.

  • Comp
    • Code <./formulas/src/Experiment.hs> [TODO ADD MISSING DEFINITIONS]
  • Fig L2 vs time K V
    • Image file: <./output/report/fig_l2_vs_time_k_v.png>
  • Fig Time Bias
    • Histograms showing the posterior sample for both algorithms.
  • LHS
    • Code: <./openmole/lhs.oms>.
  • Run Comp LHS
    • Code: <./openmole/lhs.oms>.
    • Data files: One file per pair (apmc, mon_apmc), one per simulation, in <./output/formulas/run_comp_lhs/>. Each file is the corresponding haskell value (comp, (run_apmc, run_mon_apmc)) formatted via its Show instance, such that it can be read again easily in haskell.
  • L2 Ratio
    • Code def: <./formulas/src/Experiment.hs>.
  • Time Ratio
    • Code def: <./formulas/src/Experiment.hs>.
  • Stats Comp
    • Code def: <./formulas/src/Experiment.hs>
  • Stats Comp LHS
    • Data files: <./output/formulas/figure_data/stats_comp_lhs>. A table where each line corresponds to a point in the LHS, 7 columns give the values for each parameter to comp and six additionnal columns the l2 error and runtime for each algorithm and l2 ratio and time ratio. The first line gives the columns labels.
  • Fig L2 vs biasFactor
    • Code: Gnuplot
    • Image file: <./output/report/l2_vs_bias_factor.png>.
  • Fig Scatter L2 Time LHS
    • Code: <./report/l2_time_lhs.gnuplot>
    • Image file: <./output/report/l2_time_lhs.png>.
  • Fig L2 Effects LHS
    • Code: <./report/fig_l2_effects_lhs.gnuplot>
    • Image file: <./output/report/fig_l2_effects_lhs.png>
  • Fig L2 Ratio Effects LHS
    • Code: <./report/fig_l2_ratio_effects_lhs.gnuplot>
    • Image file: <./output/report/fig_l2_ratio_effects_lhs.png>
  • Fig Time Effects LHS
    • Code: <./report/fig_time_effects_lhs.gnuplot>
    • Image file: <./output/report/fig_time_effects_lhs.png>
  • Fig Time Ratio Effects LHS
    • Code: <./report/fig_time_ratio_effects_lhs.gnuplot>
    • Image file: <./output/report/fig_time_ratio_effects_lhs.png>
  • Test Cases
    • Code: <./Makefile>
  • Stats Comp Test Cases
    • Code: <./formulas/src/Experiment.hs>
    • Data file: <./output/formulas/figure_data/stats_comp_test_cases>
  • Fig Comp Test Cases
    • Code: <./report/fig_comp_test_cases.gnuplot>
    • Image file: <./output/report/fig_comp_test_cases.png>