Quizes Flashcards

1
Q

A mathematical function f(x) has always only one global minimum. (TRUE or FALSE)

A

FALSE. Since a function can have any number of global minima or maxima. Consider, for example cos(x): This function has global minima (f = -1) at x = pi, 3pi, 5pi etc.)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Local optima occur only where f’(x) = 0.

A

FALSE. It is true that local optima occur where f’(x) = 0, but they can also occur at points where f’(x) does not exist.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

What is the feasible set (In connection with constrained optimization)? (pick one answer)

A. The set of all points for which the function reaches its optimum.
B. The set of all points that fulfil the constraints.
C. The set of all points for which f(x) = 0

A

B. The feasible set is the set of all points that fulfil the constraints.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

Which property applies to the Hessian matrix H (of a function f)? (pick one answer)

A. It is symmetric
B. Only the diagonal elements are non-zero
C. All its elements are always non-zero

A

A. The Hessian matrix is symmetric.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

Consider a 2x2 matrix in which all elements are equal to 1. Is this matrix positive definite?

A

No. The equation for the eigenvalues will be: (1-lmbda)^2 – 1 = 0, with solutions: lambda_(1,2)= 1 ± 1, i.e. 0 and 2. Thus, not all eigenvalues are positive, meaning that the matrix is not positive definite. It is, however, positive semi-definite, since all eigenvalues are non-negative.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

In an iterative method, once the search direction has been found, the problem of finding a suitable step length is one-dimensional.

A

Yes, this is TRUE. No matter the dimensionality of the vector x, once the search direction has been inserted, the resulting equation depends only on the step length (eta)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

Newton-Raphson’s method will always converge to a local minimum. (TRUE or FALSE)

A

FALSE. Depending on the starting point, the method might converge to a minimum or a
maximum. Thus, once an optimum has been reached, one must check whether it is a minimum or a maximum (for example by considering the second derivative of f(x)).

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

An optimization problem is always convex if f(x) is convex. (TRUE or FALSE)

A

FLASE. It is necessary, but not sufficient, that f(x) should be convex. In order for an optimization problem to be convex, the constraints defining the feasible set S must fulfil certain criteria. More specifically, the inequality constraints must be convex, and the equality constraints must be affine.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

Consider the level curves of a function f = f(x1,x2) and a constraint h(x1,x2) = 0. At the local optima of f (subject to h), the following holds: (pick one answer)

A. The gradients of f and h are perpendicular
B. The gradient of h is equal to the zero vector
C. The gradients of f and h are parallel

A

C. The gradients of f and h are parallel at the local optima.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

Stochastic optimization methods can be applied even if the objective function f(x) is non-differentiable. (TRUE or FALSE)

A

TRUE, and this is one, among several, advantages with such methods. Stochastic optimization methods do not explicitly make use of gradients (or higher-order derivatives), and they can therefore handle non- differentiable objective functions.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

If two individuals are able to have offspring, they must be of the same species. (TRUE or FALSE)

A

FALSE. Of course, the most common case is that the offspring is a result of the union of two individuals of the same species. However, individuals of some closely related (but different) species (e.g. lions and tigers) can have offspring. In that case, however, the offspring is not fertile, i.e. it represents an evolutionary dead end.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

Acquired traits (i.e. things learned) can be transmitted to the next generation. (TRUE or FALSE)

A

No, this is (generally) FALSE. The transferable information is stored in the germ cells (sex cells) of an individual, and does not change during the lifetime of the individual. However, it is possible that changes that do not involve alterations in the sequence of base pairs, for example changes in gene activity (i.e. gene expression) can be transmitted (a concept known as epigenetics).

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

Humans have more chromosomes than any other known species. (TRUE or FALSE)

A

FALSE. Many other species (e.g. some species of fish) have more chromosomes.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

Why must codons (sequences of bases coding for an amino acid) consist of
at least 3 letters?

A

This is so, since there are 20 (standard) amino acids. If the codon contained only two letters, they could encode at most 42=16 amino acids. With three letters, there are 64 possibilities, leading to some
redundancy.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

All processes in Darwinian evolution are random and without direction. (TRUE or FALSE)

A

FALSE. Mutations are random (undirected), where selection most certainly is not. Selection is of course not predetermined (there is a stochastic element), but the process strongly favors fit individuals.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

In an evolutionary algorithm, the fact that chromosomes are initialized randomly
slows down the process of finding the optimum. (TRUE or FALSE)

A

FALSE. In general, EAs are very fast in the early generations, quickly making up for their random starting point, normally in much shorter time than it would take for a person to assign some suitable non-random chromosomes.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

It is important to make the right choice between binary and real-number encoding,
for a given optimization problem. (TRUE or FALSE)

A

FALSE. There is no evidence that one encoding scheme provides (generally) faster convergence than the other. Often, real-number encoding is used for simplicity, at least if the number of variables is large enough (more than 10, say) to allow crossover to operate as it should.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
18
Q

In tournament selection, the number of individuals participating in a tournament
is always equal to two. (TRUE or FALSE)

A

FALSE. The size of the tournaments can be generalized to any positive integer > 2 (up to the population size), even though two is the most common tournament size. See p. 50 for a description of how to implement tournament selection in cases where tournament size is larger than two.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
19
Q

In some cases, after selecting two individuals, one does not carry out crossover. (TRUE or FALSE)

A

TRUE. Crossover, which is a very efficient operator, is only carried out with a certain probability (the crossover probability), to avoid premature convergence. In cases where crossover is not carried out, the two selected individuals (obtained, as usual, with either roulette-wheel selection or tournament selection) are subjected only to mutation before being inserted in the new
population.

20
Q

In connection with the schema theorem, what is a building block?

A

A building block is a schema with (i) low defining length, (ii) low order, and (iii) above-average fitness.

20
Q

During mutations (in GAs) each gene is considered separately, so that the number of mutations can, in principle, be anything from 0 to m (= the number of genes). (TRUE or FALSE)

A

TRUE. During mutation, one runs through each chromosome gene by gene, checking
whether or not the current gene should be mutated. Thus, in principle, the number of mutations can be anything from 0 to m, even though, typically, the mutation rate is set to a small value (usually 1/m) so that, on average, only a few mutations (or zero) occur.

21
Q

The function f(j) = j(m-j), where j is the number of ones in the chromosome,
is an example of a function of unitation. (TRUE or FALSE)

A

TRUE. Any fitness function in which the fitness depends only on the number of ones in the chromosome is a function of unitation.

22
Q

There are various methods to avoid premature convergence. However,
another good option is simply to rerun the GA a couple of times. (TRUE or FALSE)

A

TRUE. Even though it is worthwhile to implement some methods for reducing the
probability of premature convergence (e.g. a crossover probability < 1), one still often ends up rerunning the GA a couple of times. If such runs are very time-consuming, it can be a good idea first to make a few short runs in order to find suitable GA parameter values for the problem at hand, and then
to make one (or at least rather few) long runs.

23
Q

Why does one use two-point crossover in LGP?

A

Two-point crossover is used in order to achieve (the possibility of) length variation among the chromosomes. Typically, LGP is applied in situations in which the optimal size of the chromosomes is not known, a priori. Thus, it is crucial to allow length variation. Note that length variation can also be arranged with single-point crossover (if one selects different crossover points in the two chromosomes),
but it is common in LGP to exchange a middle section (thus requiring two-point crossover). With two-point crossover one can better control the magnitude of the change in the new chromosomes than with
single-point crossover: If the exchanged middle sections are not very long, the change in the chromosome will generally be smaller than with single-point crossover.

24
Q

In LGP, regarding constants, one is stuck with only the predefined ones, which
are stored in the constant registers. (TRUE or FALSE)

A

FALSE. An LGP chromosome can (and will, if needed) build temporary constants from the available parameters. Thus, for example, if the constants 1 and 2 are stored in two constant registers, the chromosome may contain an instruction that (temporarily) builds the constant 3 (e.g. r1 := c1 + c2 = 1+ 2 = 3) etc. Provided that there are enough variable registers, the temporary constants can remain in one of the variable registers for some time (possibly until the end of the evaluation, even), and can thus (once formed) be available throughout the evaluation of the LGP chromosome.

25
Q

In interactive evolutionary computation (IEC) one should try to present (on the
screen) as many solution candidates (to the user) as possible. (TRUE or FALSE)

A

FALSE. A common problem in IEC is user fatigue, from having to assess the relative merits of many candidate solutions. Typically, one therefore only presents, say, 3x3 or 4x4 solution candidates (at a time) to the user.

25
Q

Gene regulation plays a role in long-term memory and learning. (TRUE or FALSE)

A

TRUE (at least in the case of Aplysia, but in all likelihood in other organisms as well): Long-term
memory depends on changes in gene regulation that, in turn, modify synapse strength (permanently).

26
Q

Biological neural networks are able to carry out complex tasks because
neurons operate really fast. (TRUE or FALSE)

A

FALSE. A neuron needs to recover after firing a spike, meaning that the firing frequency is limited to around 1 kHz. The ability of a brain to carry out complex tasks is instead due to the parallel nature of
its computation. In the human brain, there are around e12 neurons => around e15 operations/s.

27
Q

When fitting a model (e.g. a neural network) to a given data set, one should
run the training procedure for as long as possible. (TRUE or FALSE)

A

FALSE. If the training procedure is allowed to run indefinitely, the model will fit the noise in the data, resulting in reduced quality (of interpolation, prediction etc.). Instead, one should measure the error over a separate validation set (while training), and use the error over that set (which is not fed back to the training algorithm) to determine when to stop training.

27
Q

Feedforward neural networks require backpropagation for training. (TRUE or FALSE)

A

FALSE. In general, artificial neural networks (of which feedforward neural networks is an
example) constitute a computational structure to which many different training methods can be applied. Backpropagation is often used in cases where one has a set of input-output pairs (so that an error signal
can be formed, as the difference between the desired output and the actual output, for every input). In other situations, where a set of input-output pairs is not available, other methods, such as genetic algorithms, may be used.

28
Q

What is stigmergy (pick one answer)

A. A mechanism for long-range communication using sound.
B. A mechanism for indirect communication by means of (local)
modification of the environment.
C. A mechanism for long-range communication using vision.

A

B. Stigmergy is a mechanism for indirect communication by means of (local) modification of the environment.

28
Q

Ants have skilled leaders and are capable of long-distance communication. (TRUE or FALSE)

A

FALSE. Their complex cooperative behavior results from local interactions, and is an
emergent property of the ant swarm as a whole.

28
Q

The movement of ants is always in the direction of the strongest pheromone scent. (TRUE or FALSE)

A

FALSE. Ants move probabilistically, but tend to follow the direction of the strongest scent.

28
Q

The change in the artificial pheromone level (in AS, applied to the TSP) on edge (i,j)
depends only on the (inverse of) the length of that edge. (TRUE or FALSE)

A

FALSE. The change in pheromone depends on the inverse of the length of the entire
path, 𝐷𝑘.

29
Q

In general, what is a phenomenological model?

A

A phenomenological model is a model that can describe (or reproduce) a phenomenon (such as, for example, swarming) on some level, but without providing a detailed theoretical understanding of the phenomenon. For example, the Boids model can reproduce the behavior of swarms but it does not, for example, give detailed information about the decision-making (on the level of neurons) in the brains of the birds. (Note: In addition to this definition, it should be mentioned that the term phenomenological model is also used in the field of psychology, where it carries a different meaning).

29
Q

In robot swarms (such as Harvard’s 1000-robot swarm), only local communication
(e.g. communication within the line of sight) is used. (TRUE or FALSE)

A

TRUE

29
Q

In ACO, the artificial ants always follow the edge for which the probability p(eij|S)
is highest. (TRUE or FALSE)

A

FALSE. Artificial ants select their moves probabilistically.

30
Q

In MMAS, one can prove that the probability of finding the best solution at least once in the first k iterations, tends to 1 as k tends to infinity. The reason that one
can prove this is the fact that (in MMAS) all edges have positive pheromone levels. (TRUE or FALSE)

A

TRUE

30
Q

Regarding the interpretation of the particle (individual) positions xi , what is
the main difference between PSO and GAs?

A

In PSO, there is the notion of a velocity in the search space, i.e. a smooth, continuous movement from a given position to another. This is not the case in GAs, where the movement in search space is generated by crossover and mutation.

31
Q

In PSO, both velocities and positions are restricted to a given range. (TRUE or FALSE)

A

FALSE. Only the velocities are restricted. Doing so is crucial for maintaining the coherence of the swarm, i.e. for preventing particles from being ejected from the swarm. However, there is no need to restrict positions. Indeed, the fact that positions are not restricted is an advantage of PSO (relative to GAs, for example), in the sense that, in PSO, particles can venture a bit outside the initial range (while still avoiding to leave the swarm entirely, due to the velocity restriction).

32
Q

In the simple analytical model of PSO (Appendix B.4), in which no random numbers are used (and where there is no velocity restriction), the swarm remains bounded if and only if c1 + c2 < 4. (TRUE or FALSE)

A

TRUE. However, once random numbers are included in the simple theoretical model, the swarm will not remain bounded even if c1 + c2 < 4. In a real (applied) PSO, the swarm does remain bounded, however, due to the velocity restrictions (that, again, are not included in the simple theoretical model, but always in applications!)

33
Q

How is the trade-off between exploration and exploitation handled in PSO?

A

It is handled using the inertia weight. For large values (>1) of w, exploration is favored. For values of w smaller than 1, exploitation is favored instead.

33
Q

When comparing the performance of different settings for a stochastic optimization algorithm, one should make sure to use the same number of iterations (generations) for each setting. (TRUE or FALSE)

A

FALSE. At least if one uses different population sizes. In order to get a fair comparison, one should make sure that the number of (function) evaluations is the same for all runs. Thus, for example, if
the population size is 100 in one run and the algorithm is allowed to run for 1000 generations, in a run with a population size of 1000 one should then run 100 (not 1000) generations.

34
Q

In variable truncation PSO, all variables are integers at all times. (TRUE or FALSE)

A

FALSE. In variable truncation PSO, the rounded values are only used (temporarily) when
computing the value of the objective function. At all other times, non-rounded values are used, both for
the positions and the velocities. See also the slides from the lecture.

35
Q

For the n-parity problem, one can separate the two classes (output = 0 and output = 1)
using a single hyperplane. (TRUE or FALSE)

A

FALSE. At least two hyperplanes (or lines, for the 2-parity problems) are needed.