search

NicMcPhee / XO-bias-study

Results and write-up of our genetic programming crossover bias study
MIT License
0 stars 0 forks source link

Should we play with tournament sizes? #13

Closed NicMcPhee closed 9 years ago

NicMcPhee commented 9 years ago

In the earlier runs we did, I think the tournament sizes were always 2, i.e., we were using binary tournament selection. (Am I right about this @MKDramdahl?) The default in koza.params in ECJ is tournament size of 7, which will have a substantially stronger selection pressure than binary tournament selection. The fitness of the weaker parent is likely to be poorer with binary tournament selection, and the fitness of the two parents is likely to be closer with tournament size of 7. This means that tournament size could interact in meaningful ways with crossover bias. The impact of crossover bias on my first ECJ sine regression runs, for example, seems smaller than for our earlier runs, and this could be an important of why.

As a result I'd be inclined to do at least two tournament sizes (2 and the default of 7) just to see what a difference (if anything) that makes. Thoughts?

NicMcPhee commented 9 years ago

A quick poke at the sine regression problem with tournament sizes of 2 and 7 suggests that crossover bias has a much stronger impact with tournament sizes of 2, so I definitely think we should do runs with both tournament sizes.

NicMcPhee commented 9 years ago

The first full set of results with the Sine problem make it clear that tournament size of 7 is quite different than binary tournaments, so I'm going to "call this" and say that we really have to do at least two tournament sizes.