Population-Based ES

At the start of a population-based EA, potential solutions are once again created at random. Alternatively, they may be produced by starting from user data which define roughly the search space within which a solution is sought, and then making random mutations to this to create the initial population. 

As a population of solutions now exists, two variations are opened up which we can use to select members of the next population. In both methods, the members of the child population are chosen deterministically, in other words, through a type of tournament selection in which the better of two solutions always wins. 

(p+A) Select p parents and A offspring. Mutate all individuals, then select the best p as parents for the next generation. 

A) Create A offspring from p parents and allow only the offspring to compete for survival. It is essential that A is greater than p in order to provide some form of selection pressure. All parents are replaced by offspring in this scheme. 

Much effort has been devoted to investigating methods by which convergence in the ES can be encouraged. Rechenberg proposed the one fifth success rule which, though simple conceptually and in implementation, is still effective in most cases. It suggests that if the ratio of successful mutations to all mutations is less than 0.2, the variance of the mutation operator should be decreased, while if the ratio is greater than 0.2 the variance should be increased  

---