#Parallel Tempering and the Traveling Salesman Problem ##CS205 Fall 2015 ###Jonathan Friedman, Christian Junge, Dana Katzenelson

####Directory structure:

cs205-final-project     
+-- Code
|   +-- resources
    |   +-- a280.xml.zip
    |   +-- fri26.xml.zip
    |   +-- ipython_source_files.zip
|   +-- utils     
    |   +-- __init__.py     
    |   +-- plotting.py     
    |   +-- timer.py     
    |   +-- xml_parse.py     
|   +-- ParameterSelectionPpt.py
|   +-- TimingTest.py
|   +-- annealing_helper_functions.py     
|   +-- comparison.py     
|   +-- controller.py     
|   +-- parallel_tempering.py
|   +-- plotter.py
+-- Images	
|   +-- AccuracyVsComputation.png	
|   +-- AccuracyVsIterations.png	
|   +-- AccuracyVsTime0to4sec.png	
|   +-- AccuracyVsTime20sec.png	
|   +-- AccuracyVsTime4to20sec.png
|   +-- AccuracyVsTime_200kIter.png
|   +-- Distribution.png	
|   +-- Efficiency.png
|   +-- ElapsedTimeVsComputation.png	
|   +-- PerformanceComparison1.png	
|   +-- PerformanceComparison2.png	
|   +-- PerformanceComparison3.png	
|   +-- Speed.png
+-- SavedResults
|   +-- IndividualRuns/
    |   +--[Misc files]
|   +-- dist_hist_ppt.npy
|   +-- dist_hist_sa.npy
|   +-- dist_hist_spt.npy
|   +-- hist_best_ppt.npy
|   +-- hist_best_sa.npy
|   +-- hist_best_spt.npy
|   +-- time_hist_ppt.npy
|   +-- time_hist_sa.npy
|   +-- time_hist_spt.npy
+-- .gitignore
+-- README.md     
+-- Report.ipynb

Report.ipynb is our project website. This notebook gives the background for our project, describes our implementations, interprets our results, and discusses parallelism in the context of this project.

All code for this project is located inside of the Code\ folder.

There are multiple files that can serve as the starting point for the code in this repository. To run the code, enter python <NAME_OF_FILE> into the commandline.

• comparison.py runs parallel tempering in both serial and parallel for iteration counts of 10^3, 10^4, 10^5, and 10^6 on http://www.iwr.uni-heidelberg.de/groups/comopt/software/TSPLIB95/XML-TSPLIB/instances/fri26.xml.zip (also stored in resources). Each iteration count is run 10 times. Both types of parallel tempering run with 4 systems and for the parallel version, these systems are each in a separate process. comparison.py prints the known optimum path and its length to console and then writes to file for each run the number of iterations, the length of the best path found, and the time taken.
• controller.py runs each of simulated annealing, parallel tempering in serial, and parallel tempering in parallel for iteration counts of 10^3, 10^4, and 10^5 on http://www.iwr.uni-heidelberg.de/groups/comopt/software/TSPLIB95/XML-TSPLIB/instances/fri26.xml.zip. Both types of parallel tempering run with 4 systems and for the parallel version, these systems are each in a separate process. controller.py prints to console the known optimum path and its length, as well as the best path found and its length for each iteration count of each algorithm.

There are two files that contain data that can be converted into graphs to be used in TSP. Both are located in resources/.

• fri26.xml.zip is our primary file on which we performed most of our testing. It has 26 nodes.
• a280.xml.zip is a larger file with 280 nodes. We used it to check how our results scaled to larger data sets.

Our code for parallel tempering, both in serial and in parallel, is located in parallel_tempering.py. Both functions make extensive use of the anneal_once() and related functions in annealing_helper_functions.py. The code for simulated annealing is also located in annealing_helper_functions.py

Several ipython notebooks are referenced as resources throughout the code. They are available at resources/ipython_source_files.zip.

ParameterSelectionPpt.py performs a gradient descent parameter optimization for parallel parallel tempering. The returned parameters are then used in TimingTest.py

TimingTest.py performs many runs of each of the three optimization algorithms, averages their current path and run time at each step among the many iterations, and saves these files so that they can be plotted in plotter.py

plotter.py loads saved test results from TimingTest.py and plots them. This function was used to generate all of the performance plots for the report.

Images/

Contains all of the images that are referenced in the Report. The images are loaded into the report notebook using IPython's display.Image() command.

SavedResults/

Contains the output from several runs of TimingTest.py These results are included so that the user can generate customized plots and perform analyses on the algorithms without having to re-run lengthy simulations. The IndividualRuns/ subfolder contains the results from a single run of each algorithm, in case the user wants to observe individual runs rather than averaged results over many runs. The time_hist__, hist_best__, and dist_hist__ files are averaged results from 25 runs. The time_hist__ files contain average elapsed iteration times for 50,000 iteration runs, the dist_hist__ files contain the average best found distances at each iteration for 50,000 iteration runs, and the dist_hist__ files contain the average best found distances at each iteration for 1,000,000 iteration runs.