Evolving Objects: Yet Another Evolutionary Computation Library?
2 likes22,294 views
The document presents an open-source evolutionary computation library called 'eo', which is feature-rich and designed for ease of use with various evolutionary algorithms. It emphasizes flexibility through user-defined parameters and encapsulated operators for genetic algorithms, and includes functionalities like checkpointing, statistics, and visualization. The authors also highlight ongoing development efforts aimed at enhancing its capabilities and user-friendliness.
Evolving Objects: Yet Another Evolutionary Computation Library?
- 1. Evolving Objects: Yet Another Evolutionary Computation Library? M. Keijzer1, J. J. Merelo2, G. Romero2, M. Schoenauer 3 http://eodev.sourceforge.net/ Presented at EA’01 – October 29. 2001 1-Vrije Universiteit Amsterdam 2-Universidad de Granada 3-INRIA Rocquencourt •First •Last •Go Back •Go To •Find
- 2. What is EO? An Open Source STL-based C++ EC library Don’t’ worry, this is a functional talk :-) • Paradigm-free supersedes EP, ES, GP, GA and many more! • General operators n→m • Generic operators Representation independent • User parameter definition Command-line / parameter file • Visualization, saving/restarting and many more! • Getting started “facilities” and GUI . . . EASEA •First •Last •Go Back •Go To •Find
- 3. The technical slide Re-entering code • No global variable :-) • → co-evolution, meta-evolution, subroutine use, . . . Objects, not functions STL functors • Class with single interface/method: operator() • Private data, can be passed as a whole Combination of objects (functors) • Choosing among different possibilities at run times • A combined blip is a blip •First •Last •Go Back •Go To •Find
- 4. •First •Last •Go Back •Go To •Find From EO tutorial Checkpointing: stopping criterion and statistics ¡ ¡01 1 Main CPU cost 1¡¡0¡ 0101¡¡0101¡¡0101 ¡¡¡0¡ 0¡¡1¡ "Darwinism" (stochastic or determinist) )()()()()()()()()() ( Stochastic operators: Representation dependent ¡¡¡¡¡ ¡¡¡¡¡ ¡¡¡¡ ¡¡¡¡ Mutation, ... ¡¡¡¡¡¡¡ ¡¡¡¡¡¡¡ Offspring ¥¦¡¡¥¦¡¡¡¡¡¥¦ ¥¦ ¥¦ ¥¦ ¥¦ ¥¦ ¡¡¡¦¥¡¦¥¡¡¡ Evaluation ## ## # # # Crossover, ¡¡¡¡#¡#¡#¡ ¡#¡¡#¡#¡#¡#¡# ¡#¡¡#¡¡¡¡# ¦¥¦¥¡¡¦¥¦¥¡¡¡¡¡¦¥¦¥ ¡¡¡¦¡¦¡¡¡ ¦¥¦¥ ¥ ¥ ¦¥¦¥ ¦¥¦¥ ¡#¡¡#¡¡¡¡# # # Genitors¡¡¡£¡£¡¡¡ ¡¡¡¤¡¤¡¡¡ ¤£¤ ¤ ¤ ¤£¤ ¡¡¡£¡£¡¡¡ Generation ¡¡¡¡¡¡¡ ¡¡¡¡¡¡¡ Replacement ¤£¤¡¤£¤¡¡¤¡¤¡¤£¤¡¡¤£¤ ¡¡¡¡¡¡¡ ¡¡¡¡¡¡¡ £¤£¡£¤£¡¡¡¡£¤£¡¡£¤£ £¤£ £ £ £¤£ ¡¡¡¡¡¡¡ ¡¡¡¡¡¡¡ ! ! ! !¡¡!¡¡!¡¡!¡! Selection ¡$¡$% % $% % %¡%¡¡%¡ ¡$¡¡$¡ Stop ? %¡$¡ ¡¡¡¡¡¡¡ ¢ ¢ ¢ ¢ ¢ ¡¡¢¡¡¡¢¡¡¢ Parents © ©©©©©©©©©©© Evaluation ¡¡¡¡¡¡¡ ! ! ! !¡¡!¡¡!¡¡!¡! ¡%¡¡%¡% ¡$¡¡$¡ ¡%¡¡%¡%$ %$%$ %$%$ ¡$¡¡$¡$ ¡¡¡¡¡¡¡ ¡¡¢ ¡¡¡¢ ¡¡¢ ¢ ¢ ¢ ¢ ¢ ©©©©©© © © © © © © © © © © © © ¡¡¡¡¡¡¡ ¡¡¡¡¡¡¡ ¡¡¡¡¡¡¡ !! !! !! !!¡¡!!¡¡!!¡¡!!¡!! ¡%¡¡%¡% ¡$¡¡$¡$ ¡$¡¡$¡% ¡¡¡¡$$%$ $$%$ $$%$ ¡¡ ¡¡¡ ¡¡ ¡¡ ¢¡¡¡ ¢¡¡ ¢ ¢ ¢ ¢ ¢ ¢ ©©© © ©©©©© ©©©©© © ©©©©©© © © ©©©© © ©© '¡¡'¡¡¡¡¡¡¡¡¡¡' ' ' ' ' ' ' ' ' ' ' Best individual ¡¡¡'¡'¡'¡¡¡¡¡'¡'¡ ¨¡§¡¡¡¡¡§¡ ''¡¡''¡¡¡¡¡¡¡¡¡¡'' ¡¡¡''¡''¡''¡¡¡¡¡''¡''¡ ¡¡¡¡¡¡¡¡¡¡¡¡ '' ''' ''' ''' ''' Initialisation ¨¡¡¡¨§¨¡¡¨§¨¡¡¨§¨ §§¡¨§¡¡¡¡¡¨§¡ ¨ ¨§¨ ¨§¨ ¨ § § '¡¡'¡¡¡¡¡¡¡¡¡¡' ¡¡¡¡¡¡¡¡¡¡¡¡ ' ¨¡¨¡¡§¡¡§¡¨¡§ §¡§¡¡¡¡¡§¡ ¡¡¡§¨¡¡§¨¡¡§¨ §¨ §¨ An Evolutionary Algorithm
- 5. Representation • Any structure Later templatized over the fitness • Use of STL containers vectors, lists, . . . Relieves from (most) memory management • Easy construct complex structures and of instantiated generic variation operators •First •Last •Go Back •Go To •Find
- 6. Fitness computation Any fitness type Scalar, scalar with constraints, multi-objective, . . . • The only code to write for existing representations • No useless computation Through EOs invalid flag • Trivial embedding of plain C-code with correct interface • Call of existing objects (C, Fortran) with correct interface •First •Last •Go Back •Go To •Find
- 7. Evolution Engine Selection • Popular scalar routines Roulette, ranking, tournament, . . . • Multi-objective NSGA2 Replacement: A general scheme 1. Selector among parents and children or none 2. ± Merging with parents 3. ± Selector among merged population ± Weak and strong elitism •First •Last •Go Back •Go To •Find
- 8. Variation operators • Objects, not methods Encapsulation • Standard types – Mutation Unary, 1 → 1 – Crossovers Binary, 2 → 1 “Quadratic”, 2 → 2 • General type n→m •First •Last •Go Back •Go To •Find
- 9. General operators n→m • n → 1: Orgy operators • P → 1: Global recombination Uses the whole population as parents • 0 → m: Creation of new individuals Partial restart, SDM, . . . Embedded in a breeder object that takes care of the number of offspring •First •Last •Go Back •Go To •Find
- 10. Combining operators Real-world EAs use more than one operator of each kind! EO solution: combine them Recursive structure Proportional combination • User-defined relative weights • Roulette-wheel choice • One and only one will be applied •First •Last •Go Back •Go To •Find
- 11. Combining operators (2) Sequential combination • Each operator is applied with its own probability • From none to all of them can be applied (sequentially) Supersedes standards • GA-like using pcross and pmut • ES-like global recombination + mutation •First •Last •Go Back •Go To •Find
- 12. Generic operators Bottom-up construction of variation operators for genotypes based on STL data structures vector, list, . . . Key issues • Fixed length vs variable length • Order-dependent vs order independent Base ideas for EASEA Graphical genotype builder forthcoming! •First •Last •Go Back •Go To •Find
- 13. Generic operators (2) Crossovers on vectors/lists • N-point, uniform Order- and length- preserving • Call element crossover (fixed rate, or k times) idem • Exchange with external criterion Order- and length- disturbing Mutations on vectors/lists • Call element mutation (fixed rate, or k times) Order- and length- preserving e.g. binary mutations (bit-flip and “deterministic”) • Add/remove element Length- disturbing •First •Last •Go Back •Go To •Find
- 14. Initialization • Initializer objects Can be combined . . . the proportional way • Generic for complex genotypes from base generators • Reload saved state of some previous run • Included RNG Mersenne Twister – Results independent of hardware/OS Reproducibility – Run replay For debugging! – “True” randomness Using hardware clock •First •Last •Go Back •Go To •Find
- 15. Checkpointing What you want to do every generation • Check stopping criteria is a combined stop criterion • Compute and monitor statistics Average and best, diversity, rate of feasible guys, . . . • Save current state see Initialization • Update whatever needs to be updated e.g. dynamic or adaptive population level parameters •First •Last •Go Back •Go To •Find
- 16. Statistics Computing statistics • Write the code into a stat class • Create a stat object , “add” it to the checkpoint • It will receive the current population every generation Monitoring statistics • Create a monitor (once!) Screen (text, gnuplot), file, . . . • Add the monitor to the checkpoint (once!) • Add stat objects to the monitor • Will be displayed/saved to disk every generation •First •Last •Go Back •Go To •Find
- 17. Updates Dynamic parameter: example Class Blip has a static parameter. • Derive a class NewBlip from both Blip and Update • Write the parameter update in the update method • Use NewBlip in place of Blip • “Add” that NewBlip object to the checkpoint The update method will be called every generation •First •Last •Go Back •Go To •Find
- 18. Cons • Huge compilation times Linux and g++ • CygWinT M required on MicrosoftT M systems • Getting started . . . but – A tutorial (on-going :-( Download ready-to-use snapshot at Polytechnique – Shell scripts for new representations – EASEA :-) •First •Last •Go Back •Go To •Find
- 19. On-going • Multi-objective as user-friendly as scalar CMAP et Fractales • Constraint handling methods CMAP et Fractales • Parallelization Master-slave, island model, distributed population Granada et LIFL • Port to Java within European DREAM project •First •Last •Go Back •Go To •Find
- 20. Conclusion Available representations GAs, self-adaptive ESs, GP (symbolic regression), Voronoi. Pros summary • Highly flexible • Not-so-difficult to use :-) • Complete GUI forthcoming It’s free software: Use it – and contribute! •First •Last •Go Back •Go To •Find