search

DEAP / deap

Distributed Evolutionary Algorithms in Python
http://deap.readthedocs.org/
GNU Lesser General Public License v3.0
5.77k stars 1.12k forks source link

Invalidate fitness for individuals in the history #139

Open fmder opened 8 years ago

fmder commented 8 years ago

Just after the deepcopy in History.update(), the individuals fitness should be invalidated to avoid confusion as to their validity.

jkpaulin commented 6 years ago

I attempted to apply the suggested fix, but neither invalidating fitness of the original individual, nor the copy achieved the desired result. Deleting the fitness of the copy results in no fitness values being stored in the history at all. Deleting the fitness of the original does not prevent the surprising result that we see individuals in the history with different genomes, but identical fitness values (to 15 decimal places). A minimal example:

pop = toolbox.population(n=10)
hist = tools.History()
hist.update(pop)
...
pop, logbook = algorithms.eaSimple(pop, ...)
...
for a,b in zip([hist.genealogy_history[k] for k in hist.genealogy_history],
[hist.genealogy_history[k].fitness for k in hist.genealogy_history]):
        print a,b

Example output demonstrating that the third and fourth items have different genomes, but identical fitnesses.

[1, 3, 7, 300, 0, 15, 9, 7, 5, 5, 5, 7, 2, 9, 112] (-11.923277001181246,)
[1, 3, 7, 300, 0, 15, 9, 7, 5, 5, 5, 7, 2, 9, 112] (-13.263357290736487,)
[1, 3, 7, 300, 0, 15, 9, 7, 5, 5, 5, 7, 2, 9, 112] (-13.263357290736487,)
[1, 3, 7, 44, 0, 9, 6, 4, 5, 5, 9, 7, 6, 9, 100] (-13.263357290736487,)

I have only tested this under the eaSimple algorithm.

fmder commented 6 years ago

The proposed fix implies that the individuals will not have a valid fitness in the history. An extension of the fix would require to have a decorator for the evaluation that also sets the fitness in the history.

jkpaulin commented 6 years ago

Thank you @fmder . I've worked around the problem by using the Hall of Fame functionality. Losing fitness information from History seems a shame, as the graph output is excellent, however the current behaviour caused me quite a bit of confusion, and risked invalidating my experiment - it wasn't obvious that I should not trust the fitness values from the History. I'm a beginner with DEAP so I'm afraid I don't feel able to offer a patch with your suggested enhanced fix.