agent based modelling of technological diffusion and ... · pdf fileagent based modelling of...
TRANSCRIPT
Agent based modelling of technological diffusion and
network influence
May 14, 2013
Martino Tran, D.Phil.(Oxon)
Senior Researcher, Environmental Change Institute, University of Oxford
1
Outline
Outline
1. Study motivation
2. Analytical framework & model description
3. Sample results (agent & systems behaviour)
4. Discussion & next steps
2
Techno-behavioural dynamic framework
4. Techno-behavioural
Dynamics
• Strategic decision-making and
adoption behaviour as a
mechanism for behavioural
change.
• Technology performance and
behaviour as co-evolving
coupled dynamical system.
3. Behavioural change
• Focuses on changing individual
behaviours and lifestyles
(consumption, practices , norms)
• Partly reactionary to the
conventional focus on
technological solutions without
social context.
1. Status Quo
• Extrapolation of current macro
and micro level trends in demand
and supply.
• Conservative policy deployment
leading to incremental change.
2. Technological-optimism
• Radical departure from current
trends led by accelerated
technological diffusion
• Backed by government and
industry investment.
Demand Low
Supply
High
Supply
Low
Demand High
3
Implementing the approach – general model description
Agent Behaviour
A = ƒ (P, N, Ψ)
Technology
Change
T = ƒ (Ω, λT, X)
Decision
Analysis
P = ƒ(X, β, N)
Network
Influence
N = ƒ (Q, K)
System Behaviour
S = ƒ (Φ, λ, T, P)
4
Conventional technology diffusion models – mean field approaches
𝑑𝑃(𝑡)
𝑑𝑡 = αP(t) (1 –
𝑃(𝑡)
𝐾 )
Feedback key but no other explanatory variables
d𝑛(𝑡)
d𝑡 = M− 𝑛 𝑡 p +
q
M 𝑛 𝑡 , 𝑛 0 = 0
Additional parameters important but need to disaggregate, p & q
5
Diffusion = ƒ(A, T, P, N)
ABM derivation – need to account for variation in individual behaviour and
balancing between internal and external influence
ABMs with a general binomial form*:
Prob(t) = 1 – (1 – p) * (1 – q) ^ k(t)
t(0) = no adopt
t + 1 = prob(t) > U adopt
Re-specify model to account for technology trade-off behaviour i.e. index into matrix of
options, j and influences (P, Q, K) specific to each individual, i :
Prob(t) = 1 – (1 – Pij) * (1 – Qij) ^ Kij(t)
*Source: Goldenberg & Shapira, 2009
6
= Monte Carlo simulations used for assessing unknown individual preferences (Pij )
7
Dynamic approach,
Pni θ = Lni β ƒ β θ) d β ,
+∞
−∞
Lni β = exp(β′xni)/ exp(
j
β′xni)
Accounts for heterogeneity, better approximation
Static approach,
Pni = eβ′Xni
eβ′XnjJ
j=1
HEV’s outcompete ICE’s, survey bias?
Co-evolution of technological performance and agent behaviour
• Feedback effect between technological change and agent preferences over time
• How do different feedback effects scale up to impact the larger technological system?
8
Simple assumptions made on network influence (Qij, Kij)
9
Indirect influence, Qij, based on value estimate, V of previous adopters, nk out of a local
population, n,
Qij = V(pk), pk = nk/n
Direct influence, Kij based on previous adopters out of personal contacts, wij that has
influence, yi on an individual,
Kij(t) = ∑wijyi/∑wi
Therefore the master function is,
Prob(t) = 1 – (1 – [1
R ∗ LijRr=1 (
r)]) * (1 – [nk/n]) ^ (∑wijyi/∑wi)
See Tran, 2012 for full derivation
Technology data and assumptions
Technologies Purchase
Price
(2011 USD)
Fuel Price
(USD/
160 km)
Fuel
Consumption
(L/100 km)
Performance
(Acceleration
0-100 km/hr
in seconds)
Range
(Km on 1
tank/charge)
Environment
(Annual WTW
GHG CO2-eq
emission,
metric
tonnes)
Refuelling
Availability
(%)
Petrol 16640 12.9 8.1 6.5 567 5.9 100
Diesel 20180 11.5 6.9 8.7 714 5.7 100
HEV 26155 8.45 4.7 10 862 3.9 100
PHEV 40280 7.63 4.5 12 764 4.0 30
BEV 32780 3.74 2.4 11 117 3.6 1
FC 100000 5.00 3.9 12 384 3.2 1
Sources: See Tran, M. 2012 for full references
Rewire for random and small world network influence and sensitivity analysis
Simulations P Q K
1 0.1 0.1 0.1
2 0.9 0.1 0.1
3 0.1 0.9 0.1
4 0.1 0.1 0.9
5 0.1 0.9 0.9
1
0
Captures heterogeneous behaviour, distribution and trend effects over time
0 5 10 15 20 25 300
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Time steps
Pro
babili
ty
Petrol
Diesel
HEV
PHEV
BEV
FC
A
0 5 10 15 20 25 300
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Time steps
Pro
babili
ty
Petrol
Diesel
HEV
PHEV
BEV
FC
B
0 5 10 15 20 25 300
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Time steps
Pro
babili
ty
Petrol
Diesel
HEV
PHEV
BEV
FC
C
0 5 10 15 20 25 300
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Time steps
Pro
babili
ty
Petrol
Diesel
HEV
PHEV
BEV
FC
D
A) Mass market (price, reliability), B) Early adopter (CO2 , fuel economy) C) Trade-offs (performance
vs. environment) D) Exogenous effects (refuelling infrastructure)
1
1
High variability during early phases of diffusion, better approximation to empirical evidence
Battery electric vehicle preferences: A) Mass market (price, reliability), B) Early adopter (CO2 , fuel
economy) C) Trade-offs (acceleration) D) Exogenous effects (increasing refuelling). Network effects
held constant
0 20 40 60 80 100-1
0
1
2
3
4
5
6
7
Time steps
Tota
l
y = 1.1e-006*x4 - 0.00022*x
3 + 0.014*x
2 - 0.2*x + 0.57
Data
4th degree
Norm of residuals = 6.16
A
0 20 40 60 80 1000
2
4
6
8
10
12
14
16
18
Time steps
Tota
l
y = - 4.9e-007*x4 + 8.4e-005*x
3 - 0.005*x
2
+ 0.34*x - 0.99
Data
4th degree
B
Norm of residuals = 7.9047
0 20 40 60 80 1000
1
2
3
4
5
6
7
8
Time steps
Tota
l
y = - 6.7e-007*x4 + 0.00014*x
3 - 0.01*x
2 + 0.34*x - 0.71
Data
4th degree
C
Norm of residuals = 4.5966
0 20 40 60 80 1000
5
10
15
20
25
30
Time steps
Tota
l
y = - 1.7e-007*x4 + 3.7e-005*x
3 - 0.0032*x
2 + 0.41*x - 1
Data
4th degree
D
Norm of residuals = 6.001
1
2
ABM can capture a range of outcomes – more consistent with empirical evidence showing
multiple trajectories
0 5 10 15 200
500
1000
1500
2000
2500
3000
3500
Time steps
Tota
l
LNG
A
0 5 10 15 200
1
2
3
4
5
6x 10
4
Time steps
Tota
l
ELE
B
0 5 10 15 202
4
6
8
10
12
14x 10
4
Time steps
Tota
l
CNG
C
0 5 10 15 200
1
2
3
4
5x 10
5
Time steps
Tota
l
E85
D
Cumulative adoption for A) Liquefied natural gas (LNG), B) Compressed natural gas (CNG), C) Full
battery electric, D) Ethanol-85 (E85), US 1992 – 2008 (DOE, 2008)
1
3
B
C
Accounting for evolving social network influence – clustering effects for indirect and
direct network influence
A) Increasing indirect influence (Q), direct influence held constant (K=0.1), low individual preference
(P=0.02) held constant
B) Assume personal contacts (K) reflect local population behaviour and exerts influence on individual
Combined influence can have non-linear effects on individual adoption
QA = 0.4 QB = 0.5 QC = 0.6 QD = 0.70
50
100
150
Tota
lA
KA = 0.4 KB = 0.5 KC = 0.6 KD = 0.70
100
200
300
400
500
600
Tota
l
AB
1
4
Sensitivity analysis on all parameters
100
101
102
103
100
101
102
103
Time steps
Tota
l
Parameters S1 S2 S3 S4
S5
P 0.1 0.9 0.1 0.1 0.1
Q 0.1 0.1 0.9 0.1 0.9
K 0.1 0.1 0.1 0.9 0.9
Cumulative
adoption
range
100 - 150 850 - 950 250 - 300 150 - 200 850 - 950
S1 = Ref, S2 = Pref, S3 = Indirect, S4 =Direct, S5 = Combined
• High variability in short term but clear diffusion pattern over longer term
• Combined network effects can have as much influence as individual preference
• Indirect influence can have greater effect than direct personal contacts (‘influence
through numbers effect’)
• Indicates a mechanism to change risk averse behaviour?
1
5
Discussion and next steps
Constraints:
• Lack of network data (esp. on high investment technologies e.g. energy technologies)
Potential:
• Framework has wide ranging applications for technological and behavioural interactions
• ABM can give a distribution of possible outcomes instead of locking into single trajectories ( ‘magic
number solutions’), can deal with future uncertainty rather than prediction
• How do agents balance internal strategy/preferences (profit maximization for service provision) and
external influence (networked behaviour of competitors, changing consumer preferences, macro-
level signals e.g. investment climate, policy incentives, etc.)
• How do personal (or firm-level) networks evolve with population level behaviour (larger network of
service providers)? What is the direction and magnitude of the feedback effect?
• Can individual behaviour (habits, risk aversion, preferences, etc.) be outweighed by external network
influence over time?
• What network topologies amplify behavioural signals at the individual, local and population levels?
How does agent behaviour scale up to impact the larger system e.g. disruptive technological
change?
1
6