From registration to information II Anders Ringgaard Kristensen - - PowerPoint PPT Presentation

▶
from registration to information ii
SMART_READER_LITE
LIVE PREVIEW

From registration to information II Anders Ringgaard Kristensen - - PowerPoint PPT Presentation

Oct 25, 2022 260 likes •589 views

Department of Large Animal Sciences From registration to information II Anders Ringgaard Kristensen Department of Large Animal Sciences Outline The decision making process revisited Sensors and data source Definition of concepts Value of

slide-1
SLIDE 1

From registration to information II

Anders Ringgaard Kristensen

Department of Large Animal Sciences

slide-2
SLIDE 2

Outline

The decision making process revisited Sensors and data source Definition of concepts Value of data

  • Quality of information
  • Quality of decisions

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 2

slide-3
SLIDE 3

Making decisions

Decision making is based on knowledge:

  • General knowledge: What you can read in a textbook on

animal nutrition, animal breeding, agricultural engineering etc.

  • Context specific knowledge: What relates directly to the

unique decision problem. Examples:

  • The milk yield of dairy cow No. 678 when considered for

culling.

  • The estrus status of sow No. 345 when considered for

insemination.

  • The current daily gain of the slaughter pigs in House 5

when considering whether or not to adjust protein contents of the diet.

When knowledge is represented in a form that may be used directly as basis for a decision, we call it information.

Department of Large Animal Sciences

Slide 3 Advanced Quantitative Methods in Herd Management

slide-4
SLIDE 4

Information sources

General knowledge:

  • Look in a textbook
  • Ask an expert

Context specific knowledge:

  • Obtained through registrations (observations) in the

herd:

  • Traditional registrations
  • Test day milk yield, cow 567
  • Litter size of sow 123
  • Sensor based registrations
  • Conductivity or temperature of milk from AMS
  • Accelerations of a sow from a censor node in

an ear tag

  • Computer vision (image analysis)

Department of Large Animal Sciences

Slide 4 Advanced Quantitative Methods in Herd Management

slide-5
SLIDE 5

From registrations to information

We refer to a collection (typically in a database) of registrations of the same kind as data. We don’t use data directly for decision making (huge amounts of data). Before we can use data we need to reduce it through some kind of processing. The resulting information is used for decision making (which again requires processing: optimization).

Department of Large Animal Sciences

Slide 5 Advanced Quantitative Methods in Herd Management

slide-6
SLIDE 6

The decision making process: Summary

The purpose of monitoring is to improve the decisions Processing of registrations into information is necessary Choosing the best processing is a key issue Information is a tractable representation of context specific knowledge. Monitoring is the sub-path from registration to information During this course we will follow the path from data to decision:

Department of Large Animal Sciences

Slide 6 Advanced Quantitative Methods in Herd Management

slide-7
SLIDE 7

Trends in livestock farming

Over the last decades we have seen:

  • Computers available to farmers
  • Process computers (climate control, feeding systems,

milking systems)

  • Computer networks
  • On farms
  • The internet
  • Automatic registrations by sensors
  • Improved methods for data filtering
  • State space models
  • Bayesian networks
  • Improved methods for decision support
  • Decision graphs
  • Markov decision processes (dynamic programming)
  • Improved biological understanding

How do we include this in herd management? How do we evaluate?

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 7

slide-8
SLIDE 8

PigIT: Sensors as installed in two experimental pens

Advanced Quantitative Methods in Herd Management Dias 8

Department of Large Animal Sciences

16 pens in 4 sections are monitored by sensors and cameras

slide-9
SLIDE 9

PigIT: Data infrastructure in a herd

Advanced Quantitative Methods in Herd Management Dias 9

Department of Large Animal Sciences

slide-10
SLIDE 10

PigIT: Sensor data – what does it look like?

Advanced Quantitative Methods in Herd Management Dias 10

Department of Large Animal Sciences

Water, Feed Local temp. Section: Temp. Humidity

slide-11
SLIDE 11

Data sources: Brain storm

Live weight assessment:

  • …

Heat detection

  • …

Detection/prediction of farrowing

  • …

Detection of diarrhea

  • …

Detection of mastitis

  • …

…

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 11

slide-12
SLIDE 12

Sensor types

Flowmeters Climate sensors (temperature, humidity) Pedometers Accelerometers Vision Acoustic (e.g. coughing) AMS related Sensors provide data!

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 12

slide-13
SLIDE 13

Definition

A record: λ Data: Λ = {λ1, λ1, … , λk}, Λ ∈ Rk Processing: Ψ() Information: I = Ψ(Λ), I ∈ Rm, where m << k Decision: Θ Decision strategy: I → Θ The processing of data into information typically implies a huge reduction of data. The processing is specific to the decision problem.

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 13

slide-14
SLIDE 14

Value of data (and processing)

Improving quality of information

  • How do we assess the quality of information?
  • Numerical information: Standard deviation
  • Categorical information: AUC (area under curve)

Improving quality of decisions

  • How do we assess the quality of decisions: Utility value

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 14

slide-15
SLIDE 15

Numerical information: Number of farrowings

Example from exercise later in course: A sow farmer wish to predict the number of farrowings in a given (future) week. Different kinds of data can be collected:

  • # sows inseminated
  • Historical farrowing percentage
  • Heat detection after 3 weeks
  • Pregnancy test (ultra sound)

The information requested is:

  • Expected number of farrowings

The quality of the information is:

  • The standard deviation (or variance) of the prediction

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 15

slide-16
SLIDE 16

Example from exercises: Number of farrowings

Observations (data) s Number of sows inseminated (12) 1.66 + conception rate in herd (historical data) 1.39 + heat detection after 3 weeks (2 in heat) 0.98 + heat detection quality in herd 0.81 + ultra sound scanning of 10 sows 0.47

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 16

More data increase the precision of the prediction! Refer to exercise later in course! In the table, s is the standard deviation of the prediction.

slide-17
SLIDE 17

Categorical information

Information provided only as a state: Heat detection: {In heat, Not in heat} Pregnancy test: {Pregnant, Not pregnant} Disease diagnosis: {Diseased, Not diseased} Body Condition Score: {1, 2, 3, 4, 5} State {d1, d2, … , dn}

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 17

slide-18
SLIDE 18

Categorical information

The measurement is (either literally or conceptually) a two step procedure:

  • 1. Measurement of a continuous variable y ~ N(µi, σi

2) if the

animal is in State di . The observation can be just a number

  • r an entire vectyr.
  • 2. Assigning of a state di to the measurement depending on a

set of threshold values {τ1, …, τn-1}:

  • 1. τi-1 < y ≤ τi ⇒ State di observed

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 18

slide-19
SLIDE 19

Example – pregnancy diagnosis

State, di {Not pregnant (i = 0), Pregnant (i = 1)} Measurement Hormone level, h Distributions:

  • i = 0

h ~ N(10, 12)

  • i = 1

h ~ N(13, 12) Threshold: 11 Diagnose:

  • h ≤ 11

Not pregnant

  • h > 11

Pregnant

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 19

slide-20
SLIDE 20

Overall performance of test – ROC

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 20

By varying the threshold all combinations of sensitivity and specificity along the curve can be achieved. The circle corresponds to τ = 11.5, where the sensitivity is 0.93 and the specificity is also 0.93. The performance is measured as Area Under Curve (AOC):

  • AUC = 1: Perfect method
  • AUC = 0.5: Useless method (lottery)

AUC = 0.98

slide-21
SLIDE 21

Example – pregnancy diagnosis: Less precise!

State, di {Not pregnant (i = 0), Pregnant (i = 1)} Measurement Hormone level, h Distributions:

  • i = 0

h ~ N(10, 22) - Standard deviation doubled

  • i = 1

h ~ N(13, 22) - Standard deviation doubled Threshold: 11 Diagnose:

  • h ≤ 11

Not pregnant

  • h > 11

Pregnant

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 21

slide-22
SLIDE 22

Overall performance of test – ROC – Less precise

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 22

The quality of this information is less good, because the AUC is smaller than in the other example

AUC = 0.83

slide-23
SLIDE 23

Back to PigIT: Detection of fouling and diarrhea

Data (cf. Slide 18):

  • Pen level:
  • Water consumption per hour
  • Drinking bouts per hour
  • Temperature at the corridor
  • Temperature at the resting area
  • Feed intake per day
  • Live weight per week
  • Section level
  • Temperature
  • Humidity

Information:

  • Event (fouling or diarrhea)

Processing:

  • Dynamic linear model

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 23

Jensen, D.B. 2016. Automatic learning and pattern recognition using sensor data in livestock farming. PhD thesis. Department of Large Animal Sciences. 159p.

slide-24
SLIDE 24

Value of data (measured by ROC/AUC)

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 24

Jensen, D.B. 2016. Automatic learning and pattern recognition using sensor data in livestock farming. PhD thesis. Department of Large Animal Sciences. 159p.

OBS: X-axis has been reversed!

slide-25
SLIDE 25

Improving quality of decisions – utility value

The potential positive value of data is that they lead to better information and, thus, better decisions: Compare to evaluation of a feed additive:

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 25

Feed additive

Processing Mixing

Animals Response Data

Processing (twice)

Decision Response

slide-26
SLIDE 26

Value of data

Value of feed additive: Expected income with additive – Expected income without additive = Value of feed additive . Value of data: Expected income with data – Expected income without data = Value of data .

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 26

slide-27
SLIDE 27

Value of data

Data has only value if they improve the information and, consequently, lead to better management decisions:

  • Buying/selling of animals
  • Movement of animals
  • Insemination
  • Induction of events
  • Feed ration composition
  • Feeding level
  • Observing

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 27

slide-28
SLIDE 28

Utility value of data – The two sow problem

Problem description A smallholder sow farmer has only housing capacity for one sow. Recently his sow died. He has no gilt, so he wants to buy a new sow. His neighbor has two pregnant 2nd parity sows for sale: Sow A and Sow B. Both of them will farrow two weeks from now. The price is the same. Decision problem Should the smallholder buy Sow A or Sow B from his neighbor? Initial comments The old sow is dead – nothing can change that. Everything that doesn’t depend on the decision can be ignored in the optimization (partial budgeting principle). In this case we can ignore the price of the new sow.

Dias 28 AVEPM 2013 Schwabe Symposium, Chicago

Department of Large Animal Sciences

slide-29
SLIDE 29

The two sow problem

What kind of data would the small holder farmer like to have? Litter size of first parity! Any correlation between litter size of first and second parity?

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 29

slide-30
SLIDE 30

Elements

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 30

A record: Y1i Data: Λ = {Y1A, Y1B} Processing: Predicted litter size 10.0 0.2

8.5

Information: I = {

  • ,
  • }

Decisions: Select Sow A; Select Sow B Decision strategy:

  • : Select Sow A
  • : Select Sow B
  • : Select Sow A or Select Sow B

According to Example 6.1 the value of the data is 0.28 piglets. Would it be of any value to know the litter size of just one of the sows? If yes, what would the decision strategy look like? Refer also to exercises!

slide-31
SLIDE 31

Measure errors (continued)

Pre-selection (are animals representative) Registration of interventions (e.g. treatment or culling) Missing registrations as a source of information Selection:

  • Pigs with low daily gain stay longer in the herd (exercise)
  • High yielding cows are kept longer

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 31

slide-32
SLIDE 32

Measurement errors

Registration failure – something is completely wrong Measurement precision – observation errors Indirect measure – what we observe is not exactly what we want to know Bias – systematic deviation from true value (calibrate!) Rounding errors – usually not a major problem Interval censuring – only thresholds known (date of farrowing versus time of farrowing)

Department of Large Animal Sciences

Advanced Quantitative Methods in Herd Management Slide 32