bayesian networks - qra in dnv - university of...

14
Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January 30, 2015 Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 1 / 14

Upload: lamdieu

Post on 17-May-2018

233 views

Category:

Documents


6 download

TRANSCRIPT

Page 1: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

Bayesian Networks - QRA in DNV

Sam Johnston and Amy Spicer

January 30, 2015

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 1 / 14

Page 2: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

Working Team

SAMBa: Matthew Thomas

Students: Amy Spicer, Sam Johnston

Academics: Finn Lindgren, Simon Shaw, Karim Anaya-Izquierdo

DNV: David Worthington

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 2 / 14

Page 3: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

The problem

QRA with Bayesian Networks

Proof of ConceptSensitivityConsequence modelling

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 3 / 14

Page 4: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

QRA using Bayesian Network

Is conversion easy from event trees?

What further questions does this raise?

Toy Example:

Example with 5 nodes

Further Questions:

Does this work on the scale of DNV?

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 4 / 14

Page 5: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

A Toy Example

Isolatable.section

Blowdown

Ignition.at.time.1

Ignition.at.time.2

Explosion.level

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 5 / 14

Page 6: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

QRA using Bayesian Network

Question: What happens where there is weak conditional dependence?

Consider a group of random variables X1, ...,Xn. Let Xj have valuesin Ω. Suppose that for y ∈ Ω

P[Xj = y |Xj ,1,Xj ,2, . . . ,Xj ,jk ,Xl ] ≈ P[Xj = y |Xj ,1,Xj ,2, . . . ,Xj ,jk ]

then we conclude Xj is only weakly dependent on Xl conditional onXj ,1, ...,Xj ,jk .

This corresponds to removing an arc from Xl to Xj

Kullback-Leibler divergence

DKL(P||Q) =∑y∈Ω

P(y)lnP(y)

Q(y)

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 6 / 14

Page 7: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

A Toy Example

Isolatable.section Blowdown

Ignition.at.time.1

Ignition.at.time.2

Explosion.level

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 7 / 14

Page 8: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

QRA using Bayesian Network

Further Questions:

Is this the most appropriate way to remove arcs?Which distribution do you use?What happens when you change the distribution?Continuous distibutions

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 8 / 14

Page 9: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

Sensitivity Analysis

Question: How sensitive is the Bayesian Network to change?

Introduce hyperpriors into the model?

How do small perturbations propagate through?

Further Questions:

How do we represent uncertainty?Penalised Complexity Priors

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 9 / 14

Page 10: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

Toy Example

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 10 / 14

Page 11: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

Consequence Modelling

How do we use this network in consequence modelling

Total Risk =∑x∈Ω

P(X = x)Cost(X = x)

Escape Route Graphs

Further Questions:

Can we reduce the total risk?Expectation vs. DistributionShortest path vs. safest path

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 11 / 14

Page 12: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

Toy Example

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 12 / 14

Page 13: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

DNV Example

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 13 / 14

Page 14: Bayesian Networks - QRA in DNV - University of Bathpeople.bath.ac.uk/mlt21/talks/ITT/presentation_2_ITT1.pdf · Bayesian Networks - QRA in DNV Sam Johnston and Amy Spicer January

Any Questions

Sam Johnston and Amy Spicer Bayesian Networks - QRA in DNV January 30, 2015 14 / 14