ES-1

Question 1

nondeterministic direct search methods vs deterministic direct search methods

Answer 1

one have the element of randomness one not

Question 2

Evolution striges do not make use of what property of function ?

Answer 2

ES do not use derivatives, smoothness, convexity, separability

Question 3

describe 1+1 ES

Answer 3

-one parent generate one offspring

Question 4

one fifth rule step size adaptation of 1+1 ES?

Answer 4

• one out of five offspring will gave improvements in the function value

Question 5

describe (1, λ)-ES

Answer 5

generate multiple offspring from one parent

Question 6

1+1 ES vs (1, λ)-ES ?

Answer 6

the (1,λ)-ES is less ef- ficient than the (1 + 1)- ES unless offspring can be evaluated in parallel

Question 7

describe (μ, λ)-ES

Answer 7

(μ, λ)-ES multiple parents generate multiple offspring the select best μ offspring as the new parents.

Question 8

(1, λ)-ES vs (μ, λ)-ES in sphere problem and noise problem ?

Answer 8

-the (μ, λ)-ES with μ > 1 is less efficient than the (1, λ)-ES;
-only keep the best offspring is enough.
-in the presence of noise, the
(μ, λ)-ES can be superior to the (1, λ)-ES

Question 9

What is self-adaptation step size in (μ, λ)-ES ?

Answer 9

• the step size are adapted differently.

- good strategy parameter gives good objective parameter so the candidate will survive.

Question 10

problems of self-adaptation ?

Answer 10

• adjust the mutation step size
• selection of strategy parameters is indirect and noisy
• large populations are required to adapt more than a single parameter
• self-adaptation does not perform well for the (μ/μ, λ)-ES because the efficiency is similar to 1+1-ES

Question 11

what is Recombination ?

Answer 11

• create child that is the average of the parents.

Question 12

describe (μ/μ, λ)-ES and its advantage.

Answer 12

• with μ parents, with recombination of all μ parents, either Intermediate or Weighted, and λ offspring.
• add recombination
• robust with noise

Question 13

what is Cumulative Step Size Adaptation ?

and its advantage to noise?

Answer 13

• if consecutive steps are positively correlated, then the step size should be increased
• if consecutive steps are negatively correlated, then the step size should be decreased
• is a combination of the previous steps
• works well even in the presence of noise if μ and λ are sufficiently large

Question 14

what we learned In the presence of noise about the step size ?

Answer 14

In the presence of noise, it may be useful to make large trial steps, but small search steps.

Question 15

How do evolution strategies differ from other direct search methods?

Answer 15

• nondeterministic
• use a population of candidate solutions
• is not driven by the desire to guarantee convergence to stationary points
• emphasize adaptivity
• strong invariance properties