Risk-aware policy evaluation using agent-based simulation

Risk-aware policy evaluation using agent-based simulation, Bruce Edmonds, November 2016. slide 1

Risk-aware policy evaluation using agent-based simulation

Bruce Edmonds Centre for Policy Modelling

Manchester Metropolitan University


Simple systems…

… may be complicated but behave in predictable ways, allowing them to be represented by models... •  where one can use them to numerically forecast •  where uncertainty can be analytically estimated •  where one can get rough estimates cheaply, and

better estimates with increasing investment •  which one can sensibly plan and execute

systematically •  where there is a basically one right way of doing it •  so that one can fully understand the model


However…

Even with only two bits of wood the result can be complex See video at: http://www.youtube.com/watch?v=czLIj-4suOk


The Main Point of the Talk…

…is that complex systems need to be dealt with in a different way to that of simple systems... ...not only using different techniques but also how models about complex systems are used in policy development process needs to change including moving away from prediction. •  Simulation modelling will be increasingly important

as we try to develop better policies and deal with complex and fast moving situations

•  But it can not be ‘business as usual’ – just doing better modelling with the same modeller–policy actor relationship will not work well


Structure of the (rest of the) Talk

1.  A bit about modelling context, purposes and tensions

2.  Some of the underlying assumptions and habits that need to change

3.  An eample model – A model of Domestic Water Demand

4.  An example model – Stefano Picascia’s Modelling of the Housing Rental Market

5.  Some suggestions as to ways forward


Tensions and difficulties for the modeller

Part 1


The Complexity facing Modellers

•  Many of the situations or issues we need to understand are mixtures of: technical, social, behavioural and ecological factors

•  They are not only complicated, but also unexpected outcomes can ‘emerge’ from the interaction of the actors and internal processes

•  We do not have good general models for how people behave (regardless of what economists claim)

•  How to approach using models to understand complex phenomena is not fully developed


Different modelling purposes

Models can be used for a wide variety of different purposes, and these impact upon the kind of techniques needed and its difficulties, e.g. •  Forecasting – predicting unknown (e.g. future)

situations and outcomes •  Explanation – understanding how known

outcomes might have come about •  Theoretical Exploration – understanding a

complex model by exploring some of its properties and behaviours

•  Analogy – using a model as a way of thinking about something else


Model Scope

•  The scope of a model is the conditions under which it is useful for its planned purpose

•  Whilst this is implicit and stable for many simple systems, this is not the case for many complex ones

•  Thus trying to make scope explicit is important, and these relate to model assumptions

•  A process not included in the model (and hence outside its scope) can overwhelm the results…

•  ..but in complex systems internal processes of change can also emerge, and some of these can be usefully modelled (but only in more complex ways)


Possible modelling trade-offs

•  Some desiderata for models: validity, formality, simplicity and generality

•  these are difficult to obtain simultaneously (for complex systems)

•  there is some sort of complicated trade-off between them (for each modelling exercise)

simplicity

generality

validity

formality

Analogy Solvable Mathematical Model

Data

What Policy Actors Want


A picture of modelling

wha

t is

obse

rved

or

mea

sure

d

the

mod

el

the

mod

elle

rs

the

mod

el u

sers


Assumptions and expectations from Policy Actors

Part 2


Expectations of Scientists

•  What works well with simple systems does not necessarily work well with complex ones

•  Many of the expectations of complexity scientists by policy makers and the public come from: –  What economists have claimed to be able to do –  Or how physical scientists have been able to do

•  As I hope will be clear, complex simulation modelling can usefully inform policy making

•  But these expectations can get in the way •  So we will look next at some of these expectations


The Cost-Benefit Approach

•  Basically weighing the benefits – the costs •  As if an economist had written a manual for policy

actors in how to think (i.e. as their theory states) This assumes that one can: 1.  list the main alternative options 2.  forecast the results of these 3.  put meaningful numerical values on these 4.  decide on the best one, adopt that option •  Allows policy optimisation… •  ...if it were possible


Quantification

•  Makes life much easier for policy actors – choose the one with the biggest (or smallest) number!

•  Especially when asked to justify an approach •  But can be more misleading than helpful because

it gives a false impression of accuracy •  And implicitly leads to a focus on the measurable

and that things will ‘average out’ etc. •  Was a limitation of purely mathematical

approaches, but computer simulation does not have to be focused on these aspects

•  1D quantification is often an inadequate representation of what we need to understand


Planning and Managing Modelling

•  In a simple case one can apply an approach where one carefully plans, manages and evaluates models

•  As if this was like building a bridge! •  But in complex cases complications about what

needs to be included or not requires a more iterative approach…

•  ...where models are repeatedly built for a purpose and the lessons learnt as you go along...

•  Becuase the difficulties can not be predicted in complex cases!


No gradual approximation, but scope-limited usefulness

It is often assumed that as time and effort increase the accuracy of the results improve, but this is not the case with complex systems and models Rather in order for the outcomes to be within scope enough iterative development has to occur Before this the results are worse than nothing

Time and cost

Err

or


Compartmentalism

•  That some problems can be separated into smaller sub-problems which can be modelled more simply

•  Not true in many complex cases, where the scope of modelling is dependent on having enough of the key processes represented

•  Sometimes several different modelling approaches with different (but overlapping) assumptions can be more helpful

•  Just fiddling, incrementally expanding an existing (and failing) model will probably not help here


An Example: A model of Domestic Water Demand

Part 3


Context of model

•  As part of a broader model which sought to understand the impact of climate change on the domestic demand for water in the UK

•  For the UK government and water companies •  Looked at the impact of some present and

extrapolated weather patterns under four different future economic/cultural scenarios

•  Included sophisticated statistical models for prediction of demand

•  Plus our agent-based model as a contrasting approach


Monthly Water Consumption

REL_CHNG

.88.75

.63.50

.38.25

.130.00

-.13-.25

-.38-.50

20

10

0

Std. Dev = .17 Mean = .01

N = 81.00


Relative Change in Monthly Consumption in a small village

Date

FEB 2001

SEP 2000

APR 2000

NOV 1999

JUN 1999

JAN 1999

AUG 1998

MAR 1998

OCT 1997

MAY 1997

DEC 1996

JUL 1996

FEB 1996

SEP 1995

APR 1995

NOV 1994

JUN 1994

RE

L_C

HN

G1.0

.8

.6

.4

.2

-.0

-.2

-.4

-.6


Purpose of the Model

•  Not long-term prediction •  But to begin to understand the relationship of

socially-influenced consumer behaviour to patterns of water demand

•  By producing a representational agent model amenable to fine-grained criticism

•  And hence to suggest possible interactions and outcomes


Model Structure - Overall Structure

• Activity • Frequency • Volume

Households

Policy Agent

• Temperature • Rainfall • Sunshine

Ground

Aggregate Demand


Model Structure - Microcomponents

•  Each household has a variable number of micro-components (power showers etc.): bath other_garden_watering shower hand_dishwashing washing_machine sprinkler clothes_hand_washing hand_dishwashing toilets sprinkler power_shower

•  Actions are expressed by the frequency and volume of use of each microcomponent

•  Actions-Volume-Frequency distribution in model calibrated by data from the Three Valleys


Model Structure - Household Distribution

•  Households distributed randomly on a grid •  Each household can copy from a set of

neighbours (those within a certain distance ) •  Households have different mixtures of

motivations: self, social, global •  They decide which is the neighbour most similar

to themselves – this is the one they are most likely to copy – but all neighbours have some influence

•  Depending on their evaluation of actions they might adopt that neighbour’s actions

•  Or do the action they are used to (habit) •  Or that suggested by the policy agent


An Example Social Structure (main influence only)

- Global Biased - Locally Biased - Self Biased


Household Behaviour - Endorsements

•  Action Endorsements: recentAction neighbourhoodSourced selfSourced globallySourced newAppliance bestEndorsedNeighbourSourced

•  3 Weights moderate effective strengths of neighbourhoodSourced selfSourced globallySourced endorsements and hence the bias of households

•  Can be summarised as 3 types of households influenced in different ways: global-; neighbourhood-; and self-sourced depending on the dominant weight (though this is a simplification, all three weights and factors can play a part)


History of a particular action from one agent’s point of view

Month 1: action 1330 used, endorsed as self sourced Month 2: action 1330 endorsed as recent (from personal use) and

neighbour sourced (used by agent 27) and self sourced (remembered)

Month 3: action 1330 endorsed as recent (from personal use) and neighbour sourced (agent 27 in month 2).

Month 4: action 1330 endorsed as neighbour sourced twice, used by agents 26 and 27 in month 3, also recent

Month 5: action 1330 endorsed as neighbour sourced (agent 26 in month 4), also recent

Month 6: action 1330 endorsed as neighbour sourced (agent 26 in month 5)

Month 7: replaced by action 8472 (appeared in month 5 as neighbour sourced, now endorsed 4 times, including by the most alike neighbour – agent 50)


Policy Agent - Behaviour

•  After the first month of dry conditions, suggests AFV actions to all households (reducing water usage)

•  These actions are then included in the list of those considered by the households

•  If the household’s weights predispose it, it may decide to adopt these actions

•  Some other neighbours might imitate these actions etc.

•  Others, more self-sourced may not be influenced


Number of consecutive dry months in historical scenario

0

1

2

3

4

5

6

7

8

9

J-73

J-74

J-75

J-76

J-77

J-78

J-79

J-80

J-81

J-82

J-83

J-84

J-85

J-86

J-87

J-88

J-89

J-90

J-91

J-92

J-93

J-94

J-95

J-96

J-97

Simulation Date

Num

ber o

f con

sequ

ativ

e dr

y m

onth

s


Simulated Monthly Water Consumption

REL_CHNG

.075.063

.050.037

.025.012

-.000-.013

-.025-.038

-.050

120

100

80

60

40

20

0

Std. Dev = .01 Mean = -.000

N = 325.00


Monthly Water Consumption (again)

REL_CHNG

.88.75

.63.50

.38.25

.130.00

-.13-.25

-.38-.50

20

10

0

Std. Dev = .17 Mean = .01

N = 81.00


Simulated Change in Monthly Consumption

Date

SEP 1997

APR 1996

NOV 1994

JUN 1993

JAN 1992

AUG 1990

MAR 1989

OCT 1987

MAY 1986

DEC 1984

JUL 1983

FEB 1982

SEP 1980

APR 1979

NOV 1977

JUN 1976

JAN 1975

AUG 1973

MAR 1972

OCT 1970

RE

L_C

HN

G.10

.08

.06

.04

.02

0.00

-.02

-.04

-.06


Relative Change in Monthly Consumption (again)

Date

FEB 2001

SEP 2000

APR 2000

NOV 1999

JUN 1999

JAN 1999

AUG 1998

MAR 1998

OCT 1997

MAY 1997

DEC 1996

JUL 1996

FEB 1996

SEP 1995

APR 1995

NOV 1994

JUN 1994

RE

L_C

HN

G1.0

.8

.6

.4

.2

-.0

-.2

-.4

-.6


30% Neigh. biased, historical scenario, historical innov. dates

Aggregate demand series scaled so 1973=100

0

20

40

60

80

100

120

140

160

180

200

J-73

J-74

J-75

J-76

J-77

J-78

J-79

J-80

J-81

J-82

J-83

J-84

J-85

J-86

J-87

J-88

J-89

J-90

J-91

J-92

J-93

J-94

J-95

J-96

J-97

Simulation Date

Rel

ativ

e D

eman

d


80% Neigh. biased, historical scenario, historical innov. dates


0

20

40

60

80

100

120

140

160

180

200

J-73

J-74

J-75

J-76

J-77

J-78

J-79

J-80

J-81

J-82

J-83

J-84

J-85

J-86

J-87

J-88

J-89

J-90

J-91

J-92

J-93

J-94

J-95

J-96

J-97

Simulation Date

Rel

ativ

e D

eman

d


80% Neigh. biased, medium-high scenario, historical innov. dates


0

20

40

60

80

100

120

140

160

180

200

Jan-73

Jan-74

Jan-75

Jan-76

Jan-77

Jan-78

Jan-79

Jan-80

Jan-81

Jan-82

Jan-83

Jan-84

Jan-85

Jan-86

Jan-87

Jan-88

Jan-89

Jan-90

Jan-91

Jan-92

Jan-93

Jan-94

Jan-95

Jan-96

Jan-97

Simulation Date

Rel

ativ

e D

eman

d


What did the model tell us?

•  That it is possible that social processes within communities: –  can cause a high and unpredictable variety in patterns

of demand –  can ‘lock-in’ behavioural patterns and partially ‘insulate’

them from outside influence (droughts only occasionally had a permanent affect on patterns of consumption)

•  Thus identifying and taking measures at high-usage areas at an early stage might be sensible

•  Also that the availability of new products could dominate effects from changing consumptions habits


An Example: A Model of the Rental Housing Market

Part 4


The model

•  By Stefano Picascia, an PhD student of mine, now at Sienna University, Italy

•  Is an agent-based simulation that represents both tenants and developers co-adapting

•  Is geographically based with tenants making decisions as where to move to based on location as well as quality of housing and price

•  Developers put in captial to build/rennovate housing for tenants

•  Rents are determined by the quality and prices of surrounding housing


The Manchester Case

Waves of price changes can spread Can have different outcomes each time it is run Has also been applied to London and Beirut


Average prices in a run


Different Sectors of the City in a run


What it does and does not tell us

In the model (which is the private rental sector only): •  That change is fundamentally internally driven as well

as due to outside events •  Price oscillations are endemic to the system •  That some regions of cities will be stuck as low quality

housing for long periods of time depending on the state of neighbouring areas

•  The very high price regions stay that way •  That under certain conditions sudden ‘gentrification’

may occur to some degree raising standards but maybe also displacing existing functional communities

•  For poorer districts decline is gradual and continual between any such periods


Concluding discussion and some ways forward

Part 5


From Probabilistic to Possibilistic

•  When outcomes can not be sensibly forecast… •  And especially numerically forecast… •  …where even probability zones or 90% bounds

are misleading •  Then moving to an approach that models and

understand (more of) underlying processes... •  ...in terms of the different kinds of outcome might

be much more informative •  Each outcome tagged with its own assumptions

and scopes (if they differ)


From Forecasting to Risk Analysis

•  However much one might like forecasting, often it is simply not possible…

•  ...let alone in a way such that the outcomes from different options can be compared!

•  Predicting outcomes can be more misleading than helpful •  Rather it may be more approapriate to use models for risk

analysis – finding all the ways a policy might go wrong (or right!)

•  Techniques are available to help discover and understand how endogenous processes might result in different future possibilities

•  Which can then inform the design of ‘early warning’ monitors giving the most immediate feedback to policy makers


Informing the adaptive ‘driving’ of policy

•  Complex models are no good for policy makers! •  Because they have to make decisions on grounds

they understand and know the reliability of •  They can not (and should not) delegate this to

‘experts’ and their inscrutable models •  Rather modellers should use their modelling to

understand the key emergent kinds of outcome •  To inform:

–  the consideration of these kinds of outcome –  the design of appropriate data visalisations –  the design of ‘earl warning indicators

•  …So that policy can adapt to changing trends and events as quickly and fluidly as possible


Conclusions

•  Modelling of complex phenomena is not cheap or quick and requires iterative development

•  It will not forecast the impact of potential policies or events, but can anticipate possible future outcomes in a way intuition can not

•  There will always be a ‘scope’ – a set of conditions/assumptions a model depends upon

•  But a good model can repay its investment in terms of cost and improving people’s lives many, many times over


Summary

It is no good wishing that the world or modelling is simple and trying to ‘force’ it to

be so, one has to adapt to suit reality…

…this includes how models and modelling are used by the policy process


The End

The Centre for Policy Modelling: http://cfpm.org

These slides will be available at: http://slideshare.net/BruceEdmonds

Stefano’s model of housing was developed under this project, funded by the EPSRC, grant number EP/H02171X

Social Science Aspects of Fisheries for the 21st Century – with two Icelandic partners: MATIS and the University of Iceland