Robots & Machine learning

Question 1

Properties of robot

Answer 1

Must have all of the following:

• Autonomous: based on own decisions, not teleoperated, being able to maintain itself
• exist: being able to act in the world
• sense: being able to perceive the world
• act on senses: adapt
• achieve goals: used sense d information to a goal

Question 2

What is the frame-of-reference problem?

Answer 2

Simple underlying mechanism, but complex behaviour. ! consequence is tend to overestimate complexity of mechanisms

Question 3

What is behaviour?

Answer 3

“brain” + agent environment interaction

Question 4

What is control theory?

Answer 4

the mathematical study of the properties of automated control systems

Question 5

Cybernetics

Answer 5

studying biological systems from the level of neurons, to the level of behaviour, and trying to implement similar principles in simple robots using methods from control theory.

Question 6

Artificial life

Answer 6

focuses on computational, not physical, systems that exist within a computer, not in the real world

Question 7

excitatory connections

Answer 7

stimulating

Question 8

inhibitory connections

Answer 8

weakening

Question 9

What do we need to keep in mind with embodiment?

Answer 9

• obey rules of physics (takes time, cant be at two places etc)
• needs to be aware of itself and other bodies/objects
• limitations to movements and what it can sense
• Limit on how fast the robot can move –> react

Question 10

Learning

Answer 10

acquiring new knowledge or skills and improve one’s performance

Question 11

A robot can learn about(3):

Answer 11

• itself (internal)
• its environment (maps etc)
• other robots and people

Question 12

Benefits of learning (in robotics, 3)

Answer 12

• Perform task better –> no controller is perfect
• Adapting to changes
• Simplify programming work

Question 13

Disadvantages of supervised learning(3)

Answer 13

• a trainer is needed –> less autonomous
• not learning when operating
• first training then operating phase

Question 14

What is the training info with Supervised learning?

Answer 14

The input training info is: the wanted output for each input

Question 15

What is the input for unsupervised learning and for what is it useful?

Answer 15

Input: only inputs

Useful to preprocess inputs: cluster data in meaningful clusters

Question 16

What is the training info for reinforcement learning?

Answer 16

The input is: Evaluations (rewards/penalties)

Question 17

How does a reinforcement learning system learns then?

Answer 17

• Trial and error search –> adjusts weights

Question 18

Exploration

Answer 18

Try each state to see which is optimal.

Question 19

What is exploitation?

Answer 19

Use what a robot already has learned

Question 20

What is a control policy?

Answer 20

A complete state action pair table (most often created by reinforcement learning)

Question 21

What is temporal credit assignment?

Answer 21

when in reinforcement learning after a long time the feedback, the error has to be back-propagated

Question 22

Define biomemetic

Answer 22

machines with properties similar to those of biological systems; imitate biological systems ion some way

Question 23

What is a synthetic approach?

Answer 23

Easy rules in the brain can creatie complex behaviour. It is easier to build and then increase complexity than understand what is precisely going on based on it’s behaviors.

Question 24

What is spatial credit assignment?

Answer 24

assigning credit or blame to actions taken by members of a team and not the agent itself

Question 25

What should a robot be able to do to learn from immitation?(5)

Answer 25

• To pay attention to a demonstration
• Separate relevant task from the rest
• Match observed behaviour to own behaviour
• Adjust behaviour so imitation makes sense
• Recognise and achieve goals of behaviour

Question 26

What happens with learning by demonstration

Answer 26

Putting a robot through the task enables it to map its observations and know the task.

Question 27

For which two reasons can forgetting be useful in robotics?

Answer 27

• making room for new information

- replacing incorrect information (often not wanted with NN)

Question 28

Define Putting through

Answer 28

having the learner experience the task directly, through its own sensor

Question 29

What is connectionist learning?

Answer 29

the wide variety of neural network learning algorithms.