Selection (chapter 22~28) 발표자: 임병권
Selection Multiple copies of a solution are placed in a population by deleting some solutions The essential idea is that a solution having a better fitness must have a higher probability of selection
Selective pressure If a selection operator has a large selective pressure, the population loses diversity in the population quickly => This makes premature problem A selection operator with a small selection pressure makes a slow convergence and permits enough iterations to properly search the space => Too long time
Proportional selection Map the objective function to fitness Create a probability distribution proportional to fitness Draw samples from this distribution
Fitness scaling Supersolution problem the population loses genetic diversity Almost solutions have same fitness value effect of a random selection
Roulette wheel Selection Stochastic universal sampling
Tournament selection A group of q individuals is randomly chosen from the population This group takes part in a tournament Fitness scaling 과 무관 빠르고, 병렬처리가 가능
Tournament selection #2 With incresing Tournament size q the selection pressure increases q=6~10 is recommended
Rank-based selection The rank ordering of the fitness of the individuals within the current population determines the probability of selection Ranking eliminates the need for fitness scaling
Rank-based selection #2 Linear ranking N개의 해중 i번째 순위인 해의 값 Max: 1순위 해의 값, min: 꼴지 해의 값 Max값과 min값으로 선택압을 조정 Nonlinear ranking
Boltzmann selection Boltzmann trial Competition solution i and j, in which i wins with logistic probability High T: i and j win with nearly equal prob Low T: the better of the two solutions nearly always win
Other selection method Soft brood selection Holds a tournament between members of a brood of two parents Disruptive selection Has extremely good and bas solutions. Competitive selection The fitness of a solution is determined by its interaction with other solutions.
Generation gap Refers to the amount of overlap between parent and offspring Nonoverlapping(G=1) parent solutions don’t exist in next generation Overlapping(0<G<1) Some parent solutions exist in next generation
Generation gap #2 G가 1에 가까울 수록 generational하다고 하고 G가 1/N 에 가까우면 steady-state 하다고 한다. Steady-state는 빨리 수렴시키는 경향이 있으나 premature convergence의 가능성이 더 크다.