Transcript Uncertainty and Learning in Sequential Decision

Mitigation and Adaptation Under
Uncertainty
NCAR ASP Uncertainty Colloquium
Mort Webster (Penn State)
28 July 2014
Outline of this Talk
• Overview of Policy Responses
• Overview of Mitigation Policies
• Decision-Making Under Uncertainty
• Example of Mitigation Under Uncertainty
• Mitigation and Adaptation Portfolios
Policy Response to Climate Change
• Types of Responses:
–
–
–
–
–
Mitigation: Reduce Emissions
Adaptation: Reduce Damage from Climate Change
Geoengineering: Reduce Climate Change
Energy R&D: Develop Technologies for Later
Climate Research: Improve Projections
• Key Decisions
–
–
–
–
How to allocate across types
How to allocate over time
Level of effort for each type
Who pays
Key Concept: Portfolio
• A set of investments, assets, or activities
chosen so that in combination an objective is
maximized (or minimized)
• Example: in finance, portfolio of assets is
designed to balance between maximizing
expected returns and minimizing variance
• Typically, a portfolio is adjusted over time in
response to changing information
• Main motivation is uncertainty
Mitigation
• Definition: Reduction of greenhouse gas
emissions in the future to a level lower than
“what it would have been”
• Requires a “counterfactual”
• Level of Governmental Entity
–
–
–
–
International (UN FCCC)
National / Regional (U.S., E.U., etc)
State
City?
Necessary Conditions for Emissions Reductions
• Someone, somewhere has to consume less
energy, and/or
• Someone, somewhere has to use energy from a
lower carbon source (at higher cost), and/or
• Someone, somewhere has to invest resources
to lower the cost of low-carbon energy (so that
someone will use it in the future)
• Question: How can government make
incentives for these to occur?
Policy Instruments for Mitigation
• Emissions Limits
– Cap and Trade
– Emissions Standards
• Increase cost of carbon-intensive energy
– Carbon taxes
– Remove subsidies on fossil fuels
• Incentivize/subsidize low-carbon energy
– Production tax credits
– Renewable energy standards
• R&D-Focused
– Direct R&D by Government
– R&D Investment Tax Credits
Examples of National Policies
• European Union
– Emissions Trading Scheme (ETS)
• U.S.
– No federal legislation to date
– EPA regulating through Clean Air Act
– R&D-focused policies have dominated since 1988
• Australia
– Passed carbon tax in 2011
– Repealed in 2014
Global CO2 Emissions
(Deterministic)
G lo b a l C O 2 E m issio n s (G tC )
25
N o P o lic y
CCSP 750
CCSP 650
CCSP 550
CCSP 450
20
15
10
5
0
2000
2020
2040
2060
Year
2080
2100
Global Mean Temperature Change
(Deterministic)
G lo b a l M e a n T e m p e ra tu re C h a n g e fro m 2 0 0 0
5
4
N o P o lic y
S ta b iliz e C O 2 a t 7 5 0 p p m
S ta b iliz e C O 2 a t 6 5 0 p p m
S ta b iliz e C O 2 a t 5 5 0 p p m
3
S ta b iliz e C O 2 a t 4 5 0 p p m
2
1
0
2020
2040
2060
2080
Global Mean Temperature Change
Uncertainty
U n ce rta in ty G lo b a l M e a n S u rfa ce T e m p e ra tu re C h a n g e
2 0 0 0 -2 1 0 0
1 .2
N o P o lic y
CCSP
CCSP
CCSP
CCSP
P ro b a b ility D e n s ity
1 .0
0 .8
750
650
550
450
S ta b iliza tio n
S ta b iliza tio n
S ta b iliza tio n
S ta b iliza tio n
0 .6
0 .4
0 .2
0 .0
0
2
4
6
8
D e c a d a l A ve ra g e S u rfa c e T e m p e ra tu re C h a n g e
(2 0 9 0 -2 1 0 0 ) - (2 0 1 0 -2 0 0 0 )
10
Communicating the Odds of
Temperature Change
Communicating the Impact of Policy
No Policy
Stringent Policy
(~550 ppm)
Limitations of this Example?
• Correctly states the No Policy case, BUT…
– It is NOT a once-and-for-all decision!
– What we learn along the way will change the odds.
– What we do along the way will change the odds.
– What we learn along the way will change what we
do, and the converse is also true.
Decision Analysis:
A Natural Approach
• Tools for modeling a decision problem
• Decisions under uncertainty
• Sequential decisions over time
• Effect of resolving uncertainty
• Value of information
Decision Under Uncertainty
If you are uncertain, you should:
1. Ignore it and hope it goes away?
2. Pick the most likely case (middle scenario)
and optimize for that?
3. Find best strategy for each scenario and
then average the strategies?
4. Plan for the worst case?
5. Something else?
The “Question”
Since climate change is so
uncertain, shouldn’t we just
wait until we know more?
Why Does Uncertainty Matter?
0 .3 0
If $10 million
investment is required,
would you do it?
What about $60m?
P ro b a b ility D e n sity
0 .2 5
0 .2 0
0 .1 5
m = $30m
0 .1 0
0 .0 5
0 .0 0
0
20
40
60
R e ve n u e s ($ M illio n )
80
100
Why Does Uncertainty Matter?
• If no learning is possible and no risk aversion
– Make decision based on expected value
• If you can learn and revise along the way
– May want to do more or less at first (hedging)
• If you are risk-averse
– You care about more than mean outcomes
Why Does Learning Matter?
• Why would you do something different today
if you can learn tomorrow?
• One answer: if the outcome is irreversible
Irreversibilities in Climate Change
• GHG Concentrations
– Temperature Change
– Climate Damages
• Capital Stock / Economic Investments
Example Analysis of Decision under
Uncertainty and Learning
• Focus on Climate Sensitivity
– One of the critical uncertainties
– Defn: Amount of global mean temperature change from a
doubling of CO2 at equilibrium.
– Meaning: Represents net effect of feedbacks in the
atmosphere.
• Modeling the problem
–
–
–
–
Use DICE model instead of MIT
Dynamically optimizes over time
Stabilize temperature instead of concentrations
Carbon price as measure of policy stringency
Current Uncertainty
in Climate Sensitivity
0 .3 0
P ro b a b ility D e n s ity
0 .2 5
0 .2 0
0 .1 5
0 .1 0
0 .0 5
0 .0 0
0
2
4
6
8
o
C lim a te S e n s itivity ( C )
10
Decision-Making under Uncertainty
Some Simple Starting Points:
1) What should you do if you KNOW?
2) What should you do if you will NEVER
LEARN?
3) What should you do if you don’t know, but
WILL LEARN at time T?
4) What should you do if you don’t know and
will reduce your uncertainty at time T?
Illustration: Optimal Carbon Tax
for Temperature Target of 2oC
400
C a rb o n T a x ($ /to n )
300
CS
CS
CS
CS
=
=
=
=
1 .5
3 .0
5 .0
8 .0
200
100
0
2000
2020
2040
2060
2080
2100
3 .0
o
G lo b a l M E a n S u rfa c e T e m p e ra tu re C h a n g e ( C )
Realized Temperature Change for
2 Degree Target
2 .5
CS
CS
CS
CS
=
=
=
=
1 .5
3 .0
5 .0
8 .0
2 .0
1 .5
1 .0
0 .5
0 .0
2000
2020
2040
2060
2080
2100
2120
2140
QUESTION: What should we do
now if we are uncertain?
• Wait (do nothing until we know more)?
• Implement Highest Tax (worst case)?
• Implement Lowest Tax (best case)?
• Something in the Middle?
– Where in the middle?
Relative to Best Policy in Each Case…
400
P e rfe ct In fo rm a tio n
C a rb o n T a x ($ /to n )
300
200
100
0
2000
2020
2040
2060
2080
2100
If We Learn in 2020
400
P e rfe ct In fo rm a tio n
L e a rn in 2 0 2 0
C a rb o n T a x ($ /to n )
300
200
100
0
2000
2020
2040
2060
2080
2100
Simulating Learning:
Stochastic Programming
400
C a rb o n T a x ($ /to n )
300
CS
CS
CS
CS
=
=
=
=
1 .5
3 .0
5 .0
8 .0
200
100
0
2000
2020
2040
2060
2080
2100
If We Learn in 2030
400
P e rfe ct In fo rm a tio n
L e a rn in 2 0 2 0
L e a rn in 2 0 3 0
C a rb o n T a x ($ /to n )
300
200
100
0
2000
2020
2040
2060
2080
2100
If We Learn in 2040
400
C a rb o n T a x ($ /to n )
300
P e rfe ct In fo rm a tio n
L e a rn in 2 0 2 0
L e a rn in 2 0 3 0
L e a rn in 2 0 4 0
200
100
0
2000
2020
2040
2060
2080
2100
If We Never Learn
400
P e rfe ct In fo rm a tio n
L e a rn in 2 0 2 0
L e a rn in 2 0 3 0
L e a rn in 2 0 4 0
N e ve r L e a rn
C a rb o n T a x ($ /to n )
300
200
100
0
2000
2020
2040
2060
2080
2100
Summary – Effect of Learning Later
with 2o target
100
O p tim a l T a x in 2 0 1 5
90
80
70
60
50
40
30
2020
2030
2040
2050
2060
T im e W h e n L e a rn in g O ccu rs
n e ve r
Decision Under Uncertainty
With Partial Learning
400
P e rfe c t In fo rm a tio n
a fte r 2 1 0 0
C a rb o n T a x ($ /to n )
300
R e vis e d p o lic y a fte r
re d u c in g u n c e rta in ty
200
N e a r-te rm h e d g in g
u n d e r u n c e rta in ty
100
0
2000
2020
2040
2060
2080
2100
2120
2140
Main Point of this Example
• If I am uncertain, but can learn and revise later
• The best decision for today is (almost)
NEVER:
– Do “nothing”
– Assume the “worst-case”
– Do the “average” decision from a range of
scenarios
• The optimal “hedge” depends on specific
characteristics of your problem
Methods for DMUU
• Decision Analysis
• Dynamic Programming
• Stochastic Programming
Essence of Dynamic Programming
Vt s = max 𝐶 𝑥, 𝑠 + 𝐸 𝑉𝑡+1 (𝑥, 𝑠)
𝑥
• Words: Balance between immediate reward
and expected value of being in a better position
next time
Mitigation and Adaptation
• Increasing prominence in policy discussions
and in research
• Early Treatment of M vs A:
–
–
–
–
–
Substitutes or complements?
Framed as a static decision
Framed as a deterministic question (no uncertainty)
Framed around a generic “policymaker”
“Adaptation” as a monolithic concept
Realities Surrounding M vs A
• Decision-Makers:
– Different decisions are made at different levels
– Local DMs: can only manage specific adaptation
investments
– National/Regional: Allocation of resources across
general categories
• Relative Allocation over Time
– Should be a portfolio that is adjusted over time in
response to changing conditions and information
Realities Surrounding M vs A II
• Substitutes vs. Complements?
– Mitigation today substitutes for future adaptation
– Adaptation today is for climate change already
occurring
• Adaptation Types
– “Flow” Adaptation – requires repeated application
– “Stock” Adaptation – long-lived investments
– “Option” Stock Adapt. – Build in additional
flexibility to infrastructure
Mitigation vs Adaptation
• Effective response will be a portfolio over
time of mitigation and adaptation investments
• Precise mix will evolve with new information
• Today’s mix should be informed by
– Level of near-term climate damages expected
– Uncertainties in future climate damage, mitigation
costs, and adaptation costs and effectiveness
– Expected time to obtain information on the
uncertainties
Summary
• Climate policy decision is a risk management
problem
• Response should be a portfolio of policies
designed to reduce risk as well as minimize
expected damages
• Near-term response will hedge against the
uncertainty
• For research, appropriate methods include
decision analysis, dynamic programming, and
stochastic programming