Hyperparameter Optimization Flashcards

1
Q

(True or false) hyperparameter optimization (HPO) is a black-box optimization problem.

A

True

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

R package that provides different tuning approaches

A

mlr3 tuning

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

The relevant parameter set can be constructed / specified using the ___ command.

A

ps()

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

All the details for the tuning procedure need to be combined in a ____ .

A

Tuning instance

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

It specifies the method that is used for optimizing the tuning instance.

A

Tuner

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

It defines the number of (equidistant) values per configurable parameter that are to be tried during the tuning procedure.

A

Resolution

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

To avoid overfitting, tuning itself should be performed during the ___.

A

Training procedure

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

A mixture of learner and tuner

A

Autotuner

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

(True or false) “irace” returns the elite configurations and can be used for tuning.

A

True

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

Finding a good hyperparameter configuration for
the problem at hand.

A

Hyperparameter optimization

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

Which are the two main types of baseline optimizers?

A

Grid search and random search

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

Baseline in which a number of values is selected for each parameter and evaluate all possible combinations.

A

Grid search

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

Baseline in which each parameter is sampled uniformly at random.

A

Random search

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

Evaluating novel configurations (very) different to previous ones.

A

Exploration

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

Two types of stochastic approaches for hyperparameter optimization.

A

Simulated Annealing and CMA-ES

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

Two types of model-based approaches for hyperparameter optimization.

A

iterated F-race (irace) and Hyperband.

17
Q

Trying to improve existing configurations by evaluating similar ones.

A

Exploitation

18
Q

True or false? In the context of Bayesian optimization, the „expected improvement“ acquisition function trades off exploration and exploitation of the search space. Therefore, it is a suitable method in situations where functions are expensive to evaluate, e.g. hyperparameter configurations in large-scale problems.

A

True

19
Q

Another name for Bayesian Optimization

A

Sequential Model-Based Optimization, SMBO

20
Q

A fast-to-evaluate model of performance function, based on
already evaluated configurations, with uncertainty estimate.

A

Surrogate model

21
Q

Optimization which evaluates more configurations with lower budgets speeds up HPO by terminating bad configurations early, and thus allows better optimization with limited overall budget.

A

Multifidelity optimization

22
Q

Type of multifidelity optimization that starts all configurations with a certain fraction of the budget. Then, discards half of the configurations with worst performance, doubles the budget and repeats until final budget is reached.

A

Succesive Halving (SH)

23
Q

The three most important parameters of succesive halving.

A

Budget factor, final budget, and total used budget.

24
Q

This algorithm tries to alleviate the problem of how
well the initial training phase of SH captures final performance by introducing multiple brackets.

A

Hyperband

25
Q

(True or false) The smallest allocated budget increases with each consecutive bracket in the hyperband algorithm.

A

True

26
Q

Another advanced tool for (multifidelity) algorithm configuration which uses statistical tests to identify and stop bad configurations. Its central concept is “racing”.

A

irace

27
Q

(True or false) In Successive Halving, the budget allocated to the runs is halved after each iteration.

A

False. After each stage, the budget is multiplied by 𝜂 and the number of active configurations divided by 𝜂.

28
Q

True or false? Hyperparameter optimization can be understood as a special case of the CASH problem.

A

True