Prioritized Grammar Enumeration for HeuristicLab

[gkronber]

It would be great to be able to use PGE (prioritized grammar enumeration) within HeuristicLab.

The original implementation of the algorithm is in Go. So we either need to convert the implementation to C# or provide a wrapper that calls the native binaries via DllImport.

This should be done within an undergraduate thesis.

• PGE paper: ​https://github.com/verdverm/go-pge/raw/master/pge_gecco2013.pdf
• PGE implementation: ​https://github.com/verdverm/go-pge

[gkronber]

r15990: created branch

[gkronber]

r15991: prepared plugin and empty algorithm class for PGE implementation

[hmaislin]

r16079: added testfile

[hmaislin]

r16080: initial commit of working PGE version

[gkronber]

r16081: added a script to start pge with command line parameters

[gkronber]

r16182: Added PGE dll and the first working call

[hmaislin]

r16183: Adapted PGE Go code to work as dll

[hmaislin]

r16191: Added current DLL to PGE dir that needs to be placed in bin path

[gkronber]

r16195: copied go-pge.dll to plugin folder and added to project file so that the native .dll is also copied to the bin folder. Set unsafe property for the project.

[gkronber]

r16196: introduced parameter names to prepare for renaming parameters

[gkronber]

r16197: renaming things

[gkronber]

r16199: renaming things, use data from regression problem instead of hard-coded data. Algorithm crashes in StepW.

[hmaislin]

r16200: Updated DLL to get coeff results and release memory

[gkronber]

r16212: several fixes

[hmaislin]

r16229: Updated PGE DLL

[hmaislin]

r16228: Adapted pge plugin to check for null value

[hmaislin]

r16231: Fixed result error

[gkronber]

r16315:

• worked on display of PGE results.
• convert PGE expressions into HL solutions

Algorithm does not produce reasonable results (e.g. for Korns-1). I suspect that the index of the target variable is passed incorrectly (searchVar parameter in InitProblem). However, when I change the searchVar value from 0 to 1 the program crashes because of a general protection fault in the call of InitProblem

[gkronber]

r16316: fixed program crash (searchVar must be zero!)

[gkronber]

r16335: changed criterion for updating the solution to testScore < bestTestScore

[hmaislin]

r16510: Updated DLL, result error seems to be fixed, more testing is necessary

[gkronber]

r16515: changed code to update best solutions to work with empty results.

[gkronber]

I made a quick test with:

• Keijzer Benchmark Problems -> Keijzer 1
• Various Benchmark Problems -> Spatial Coevolution

With the default parameter settings the algorithm produces a list of empty results for Keijzer 1.

For the Spatial Coevolution problem I changed

• Nodes: Add Mul Div Exp
• NonTrig: Add Mul Div Exp

The log contains a few equations.

The quality of these solutions seems to be very bad.

.

Another observation is that the testScore produced by GetStepResult is always the number of rows in the dataset (testScore never improves).

[hmaislin]

rt16533: Added Powershell testscript to compare results

[hmaislin]

r16614: Fixed result bug, however testScore problem not solved yet, added multitest C and Powershell script

[hmaislin]

r16615: Added updated go and plugin files

[hmaislin]

r16617: Added missing levmar dependency libs amd64 and fixed MultiTest bugs

[hmaislin]

r16618: Added C and Go benchmark logs, fixed some MultiTest problems

[hmaislin]

r16620: Reorganized folder structure for make script, removed explicit marshalling, erased go-side logging

[gkronber]

r16621: changed code to update the best found solution

[hmaislin]

r16666: Modified make / folder structure to build original app, added x86 dll

[hmaislin]

r16669: Added missing dependencies