Answers

U12.1 Learning a behaviour

Learning, and the capstone · University · about 30 min

What this lesson is about

What learning buys over a hand-written rule, what it costs, and the smallest honest example.

• What learning is, exactly
• Three settings you will meet
• When learning is worth it
• When it is not
• About position() in this module
• The smallest example there is

Questions 7 marks in all

1. [1 mark]Which three things must be present for there to be anything to learn?

   Tick every answer that is true.

   1. AA family of candidate answers, the hypothesis class
   2. BA score: a loss to make small or a reward to make large
   3. CA search that finds a good candidate
   4. DA neural network
   5. EThousands of labelled examples
   Answer: A, B, C. Driving a leg, measuring it and dividing has all three. Networks and big data sets are one way to fill the slots, not requirements.

2. [1 mark]The data sheet says 20 cm/s at command 100. The robot drives at command 60 for 4 s and travels 44.2 cm. What does this print?

   nominal = 20.0 * 60 / 100.0 * 4.0
   measured = 44.2
   print(nominal, measured, round(measured / nominal, 3))

   Answer:

   48.0 44.2 0.921

   The data sheet predicts 12 cm/s for 4 s, 48 cm. The robot's gain is 44.2 / 48 = 0.921.

3. [1 mark]What is the name for the family of candidate answers a learner chooses between, such as all straight lines with a slope and an intercept?

   Answer: hypothesis class. Choosing the hypothesis class is a modelling decision you make, not something the data decides.

4. [1 mark]A robot tries lateral offsets on short trials and keeps whichever scores best. Which setting is this?

   1. APolicy search: the data is trials and their scores
   2. BSupervised learning: the data is inputs with the right answers attached
   3. CSystem identification: the data is commands and what the robot did
   4. DNone, because no model is fitted
   Answer: A. Nobody can tell the robot the right offset, only how well a trial went, so the search has to be done by trying.

5. [1 mark]Which of these is a poor use of learning?

   1. ALearning the inverse kinematics from data when the geometry gives them exactly
   2. BMeasuring this robot's actual speed per unit of command
   3. CRefitting the drive model after changing to a new mat
   4. DLearning what 64 depth readings mean about which way is passable
   Answer: A. When you have the equation, learning it wastes data and does worse than the formula. The others are measurable but not derivable, or they change.

6. [1 mark]In this module position() is available. What is the rule for using it?

   1. AThe truth may inform the score, never the policy
   2. BIt may be used anywhere, because it is the lab's measuring rig
   3. CIt may be used by the policy but not by the score
   4. DIt must not be used at all
   Answer: A. Ground truth labels and scores training, as motion capture does in a real lab. A policy that reads it could never be deployed.

7. [1 mark]The single gain from one 4 s run predicts the robot's travel better than the data sheet. What is still missing?

   1. AAny idea of how much to trust it: one run gives no spread
   2. BA loss function
   3. CA hypothesis class
   4. DA search
   Answer: A. The run has a hypothesis class, a loss and a search. Repeating it and reporting the mean and spread is what says how far to believe it.

The task: the data sheet is not your robot

from bugbot import *
connect()

CMD = 60
SECONDS = 4.0
V_MAX = 20.0

from bugbot import *
connect()

CMD = 60
SECONDS = 4.0
V_MAX = 20.0                       # the data sheet: cm/s at command 100

nominal = V_MAX * CMD / 100.0 * SECONDS
print("nominal:", round(nominal, 1))

# the lab's overhead camera, which the robot itself does not have
x0, y0 = position()
forward(CMD)
wait(SECONDS)
stop()
wait(0.5)
x1, y1 = position()
measured = ((x1 - x0) ** 2 + (y1 - y0) ** 2) ** 0.5
print("measured:", round(measured, 1))
print("gain:", round(measured / nominal, 3))