Modelling evolving patterns of land use:The FEARLUS model

Gary Polhill and Nick GottsMacaulay Land Use Research Institute

ESRC Seminar SeriesMicrosimulation modelling in the UK: bridging the gaps

Seminar 2: ‘Adding behaviour’

Acknowledgements

• Various collaborators:– Alistair Law, Dawn Parker, Luis Izquierdo,

Lee-Ann Sutherland, Pernette Belveze, Pete Edwards,Alun Preece, Edoardo Pignotti, Alessandro Gimona

• Funding:– Scottish Government Rural and Environmental Research

and Analysis Directorate– ESRC National Centre for eSocial Science– EU FP6 New and Emerging Science and Technology

Pathfinder Initiative on Tackling Complexity in Science

Work with FEARLUS

• The (two) original FEARLUS projects– Building and developing the FEARLUS model

• FEARLUS-CAMEL– Linking land use change to diffuse pollution

• FEARLUS-SPOMM– Linking land use change to biodiversity

• FEARLUS-ELMM– Improving the land market model

• CAVES (FP6)– Using qualitative research to enhance and validate FEARLUS– Exploring complex dynamics in FEARLUS

• Other work not (directly) relevant to land use:– ESRC funded work on enabling ABMs on the semantic grid

• Managing outputs from ABMs and making explicit experimental procedure

– Issues of numerics in ABMs• Behaviour of ABMs can be influenced by improperly handled floating-

point exceptions

Outline

• Inspiration and evidence for FEARLUS• Heuristic decision making in FEARLUS

– Early work with FEARLUS

• More sophisticated decision-making– Adding Case-Based Reasoning

• Other areas of decision-making and concluding points

Outline

• Inspiration and evidence for FEARLUS• Heuristic decision making in FEARLUS• More sophisticated decision-making• Other areas of decision-making and

concluding points

Predicting forestry in Scotland

• Forestry is profitable

• Probability of forestry– Dark green = higher P

• Based on suitability– Climate

– Gradient

– Soil type

(Aspinall & Birnie, unpub.)

Predicted versus actual forestry

• Yellow shows actual forestry

• Green shows predictions

• Large areas of high suitability but no forestry

Influence of land ownership

• Red lines show ownership boundaries

• Land use is based on more than suitability and (simple) economics

• Sociological factors– e.g. Grouse shooting

• Landscape pattern at the regional scale is a function of local interactions and individual preferences

‘Anecdotal’ evidence

• The Guardian, UK, Tuesday 24 April 2007Kemble Farms has been getting 19p per litre for milk

[a loss of ~2p per litre]

‘The irony for Colin Rank, one of the family that owns Kemble Farms, is that his cows drink water from a Cotswold spring that he could bottle and sell for 80p a litre. “We’re giving it to cows and devaluing it by turning it into milk. Like all dairy farmers we could pack up tomorrow and do something better with our capital, but we do it because we have an emotional investment in the land and the animals. And we know there’s a market for our product, if only the market worked.”’

[Felicity Lawrence]

Decision-making in rural systems

• Land manager decision making isn’t (entirely) fiscally rational– Multidimensional

• Desire to be seen by peers as a ‘good farmer’ [Burton, 2004]• Keeping the name on the farm; identity [Burton and Wilson, 2006]• Conservation

– Uncertainty means utility maximisation is not appropriate• Satisficing, heuristic strategies popular in models [Parker et al. 2008]• Can also use algorithms grounded in cognitive theory

– Interactive decision-making• Social influences: imitation, advice, approval, …

• If it were, then competition with less regulated global agricultural systems could mean bad news for:– Soil, water, biodiversity, wildlife, animal welfare, landscape amenity,

workers’ rights, food security, disease control, …

Outline

• Inspiration and evidence for FEARLUS• Heuristic decision making in FEARLUS• More sophisticated decision-making• Other areas of decision-making and

concluding points

Original model

YearlyYearlyCycleCycle

Calculation of economic return from

each land parcel

Land use selection

£

Externalconditions(last n years)

Biophysical properties

Returns and biophysical properties

from other land parcels belonging to self and

neighbours(last n years)

Biophysical properties

£

Externalconditions

Returns(last n years)

Land use

Calculationof Return

Land sales

Before

After

FEARLUS Agents

• Decision algorithm chooses land uses

• Wealth

• No theoretical limit to age provided that wealth >= 0

• Forced to buy land parcels if sufficient wealthForced to sell land parcels if wealth < 0

• Social neighbourhood

(von Neumann neighbourhood)

Decision algorithm

Satisfied with

yield?

Copy neighbour

s?

Yes

Yes

No

No

Habit

Yield Copying

Majority Copying

Optimum Copying

Random

Parcel Matching

Experimentation and Imitation

Year 12

Year 13

Year 14

Year 15

Year 16

Year 17

Some of the sub-populations

• Sub-population “SI”– Always use Majority Copying strategy

• Imitate based on a weighted selection of the number of times Land Uses appear in the neighbourhood

• Sub-population “II”– Always use Optimum Copying strategy

• Choose the best performing Land Use in the neighbourhood, with assessment based on a knowledge of differences in Biophysical Characteristics between neighbouring Parcels and that for which a Land Use is being chosen

• Sub-population “HRYI”– Use Habit strategy if yield of parcel > 11

– Else use Random strategy with probability 1/16

– Else use Yield Copying strategy• Imitate based on a weighted selection of the total Yield each Land Use

has in the neighbourhood

Comparing decision algorithms

• 30 trials containing two sub-populations– Each sub-population contains agents with a particular

decision algorithm

• The “winning” sub-population in each trial is that owning the majority of the land parcels after year 200– Binomial test used to see if one sub-population beats the

other a significant number of times

Non-transitivity of winners

II

HRYI

SI

Why didn’t II beat SI?

SI vs HRYI

• Varying dominance of land use

• SI slow on the uptake; waits for new land uses to become established

• HRYI (SubPop 2) beats SI

II vs HRYI

• Dominant land use changes over time with fluctuations in climate/market

• Steeper adoption curve

• II is able to exploit risk taking strategy of HRYI

• II (SubPop 1) ends up with more land parcels

II vs SI

• Lock-in on land use 3• SI and II both have

purely imitative decision algorithms

• Neither SI nor II is able to dominate

Wider questions on imitation

• When to copy?– When aspiration threshold breached?

• At what level should the aspiration be set?Gotts and Polhill, 2003

• There are various ways that imitation could be implemented:– Copy the highest yield in the neighbourhood– Copy the most used in the neighbourhood– (Compromise) Copy by total yield– Weighted selection from the neighbourhood

Forthcoming paper in JASSS

FEARLUS and microsimulation

• We have used some microsimulation-like studies to examine simple models– Why using an aspiration threshold (‘satisficing’) is

an advantage– Why diversifying land use selections can be an

advantage in highly unpredictable environments

• Use decision-making algorithms that don’t involve interactions with neighbours– Well suited to simple heuristic algorithms

• Use only a small number of land parcels

Possible Land Parcel Histories in Simplified Model: Random Choice vs Solvency Threshold HR Managers

Year ZeroYear Zero

Random ChoiceManager

Yield aboveSolvency Threshold

Random ChoiceManager

Yield belowSolvency Threshold

Solvency ThresholdHR ManagerYield above

Solvency Threshold

Solvency ThresholdHR ManagerYield below

Solvency Threshold

Random ChoiceManager

Yield above Solvency Threshold

Solvency ThresholdHR ManagerYield above

Solvency Threshold

Random ChoiceManager

Yield below Solvency Threshold

Solvency ThresholdHR ManagerYield below

Solvency Threshold

Year ZeroYear Zero

The Advantages of Diversity when all Land Uses are equal:Within-Estate Diversity in a Two-Parcel Environment

• Both Parcels the same.• Two possible Land Uses, Yields of BET+1/2 and BET-1/2. • Equal probability each will be the “good” Land Use in any Year.• LPP 1.

Symmetrical Random Walk model

Outline

• Inspiration and evidence for FEARLUS• Heuristic decision making in FEARLUS• More sophisticated decision-making• Other areas of decision-making and

concluding points

FEARLUS & Diffuse pollution

• Diffuse agricultural pollution generally refers to runoff from fields:– Nitrates, Phosphates, Pesticides, Faecal coliforms

• Could also be applied to airborne pollutants:– Pesticides, Greenhouse gases (methane, nitrous oxide)

• Pollutants in runoff to rivers can be monitored downstream• Increasing ability to monitor airborne pollutants, but not at

single-farm scale• Monitoring is less costly (in some cases only possible) at above

farm scale – e.g. over a catchment or sub-catchment• Could social interactions between farmers be used to make

such monitoring effective in reducing pollution?

Farmers and their Neighbours

• Farmers both compete and cooperate with peers and neighbours

• Farmers learn from their peers and neighbours• Farmers are not straightforward profit-maximisers:

they value the good opinion of peers and neighbours– In itself: to be seen as a “good farmer”– Because they may need neighbours’ help

• Refocusing: Under what circumstances would collectively-earned payment for pollution reduction be a policy instrument worth considering?

Estimated Yield

Land Uses

Land use selection

£Climate

Market Conditions

Calculation of Return

Before

Social Interactions

YearlyYearlyCycleCycle

Land sales

Neighbours’ Approval/Disapproval

Land use

Biophysical properties

PollutionEstimated Social

Acceptability

After

FEARLUS-W

Return

FEARLUS-W:Model of Farmer Decision-Making

• Profit, Social Approval and Salience– FEARLUS-W Land Managers choose land uses on the basis of expected

profit and expected approval from neighbours– Relative importance of these varies between Land Managers and over time– Change over time due to salience-changing events, e.g. a bad harvest, a

neighbour’s disapproval

• Case-Based Reasoning– Land Managers maintain an episodic memory, or case base– Every Year they consider whether they are satisfied with how their

neighbours assess them, and for each land parcel, whether it is yielding a satisfactory return.

– If satisfied, they do not change land use for that parcel.– Otherwise, they consider past experiences with each land use, selecting the

case most similar to the current case. Judgement of similarity is based on:• climatic conditions of the case compared to those expected in the coming year,

• economic conditions of the case compared to those expected in the coming year,

• proximity of the land parcel in the case to that now being considered.

– If no suitable case is found, default values are used.

Basis for Case Based Reasoning(From Izquierdo 2008, PhD Thesis)

• Case-Based Reasoning arose from Cognitive Science in the late 1970s– Knowledge gained from experience is encoded in episodic

memory as “scripts” allowing us to set up expectations and inferences (Schank & Abelson)

• Supported by psychological studies– Klein and Calderwood (1988) conclude from a study of 400

decisions that: “processes involved in retrieving and comparing prior cases are far more important in naturalistic decision making than are the application of abstract principles, rules or conscious deliberation between alternatives”

Estimated Profit

Estimated Social Acceptability

Land UsesPareto Front

Multidimensional decision making

Salience determines choice

Description of FEARLUS Runs Performed

• Toroidal environment of 20 × 20 land parcels• Spatially variable biophysical conditions and temporally variable (but auto-

correlated) climatic conditions determining yield, temporally variable but auto-correlated economic conditions then determining economic return jointly with yield.

• Each Land Manager initially owning 1 parcel (unsuccessful Land Managers sell up)

• Five land uses, with mean yield varying linearly with pollution generated (both expressed in arbitrary units) thus:

97531Pollution

11.510.59.58.57.5Mean Yield

LU5LU4LU3LU2LU 1Land Use

Description of FEARLUS Runs Performed

• Six reward conditions:– Threshold 2000, reward per land parcel 50– Threshold 2000, reward per land parcel 25– Threshold 1750, reward per land parcel 50– Threshold 1750, reward per land parcel 25– Threshold 1500, reward per land parcel 50

– No reward• Six neighbour-(dis)approval and (dis)approval salience-

increasing conditions– Base (dis)approval on absolute pollution levels, increase salience of neighbours’ opinion when

disapproved of– Base (dis)approval on absolute pollution levels, increase salience of neighbours’ opinion when

reward not given– Base (dis)approval on relative pollution levels, increase salience of neighbours’ opinion when

disapproved of– Base (dis)approval on relative pollution levels, increase salience of neighbours’ opinion when reward

not given

– Base (dis)approval on relative pollution levels, disapprove more strongly than approve, increase salience of neighbours’ opinion when disapproved of

– No concern with neighbours, so no Social Approval Function

Pollution

No Reward No Reward

Reward 50 @ 2000

No Social Approval

No Social Approval

With Social Approval

Reward 50 @ 2000

With Social Approval

Land Uses

No Reward No Reward

Reward 50 @ 2000

No Social Approval

No Social Approval

With Social Approval

Reward 50 @ 2000

With Social Approval

Decision Making Mode

No Reward No Reward

Reward 50 @ 2000

No Social Approval

No Social Approval

With Social Approval

Reward 50 @ 2000

With Social Approval

Some observations

• The larger the reward, the less time Land Managers spent using more polluting land uses

• Social approval lowered pollution– …if Land Managers cared about neighbours’ opinions and

disapproved of neighbours polluting• Taken together, the above were more effective than

either acting separately

• The general pattern of a simulation run in which the reward had an effect was of variation in pollution levels as exogenous factors changed land uses’ profitability– Levels much above the threshold were seldom maintained for

long periods• However, the algorithm used did make a difference to

the size of the reduction effect

Outline

• Inspiration and evidence for FEARLUS• Heuristic decision making in FEARLUS• More sophisticated decision-making• Other areas of decision-making and

concluding points

Other areas of decision-making in FEARLUS

• Advice model– Allows agents to exchange cases when they don’t have

experience– Used in recent work with Alessandro Gimona, on

biodiversity• Not yet adequately explored

• Land market decisions (with Dawn Parker)– Using non-optimising decision making means various

questions have to be answered:• When to sell? When to buy?• What to sell? What to buy?• What price to accept? What price to offer? What is the final sale

price?– Decisions about these affect outcomes in the model

Concluding points

• Agent-Based Modelling does not constrain assumptions about decision-making to– Optimisation/Maximisation– (Fiscal) Rationality– Non-interactive

• Decision-making in ABM can be based on cognitive theory• Results can be affected by different algorithms used to

implement behaviour– In FEARLUS, we explore alternatives

• Assumption of optimisation/rationality/non-interactivity is still an assumption– Mathematical tractability is no longer an excuse for failing to look

formally at other assumptions and algorithms