1. Economic Policy Simulation and Optimization
Peter Le
Computer Systems Research
Period 2
5/28/2009

2. Purpose
Create feasible and simple economic (taxation and welfare) model
Implement optimization algorithm effectively
Help improve public policy through test runs and simulation data

3. Economic Policy Government regulation Citizen feedback
Changes depending on demographics and economy

4. Simulation Government/Citizen relationship over a 12 year cycle
Citizen objects consume, produce, spend, and are taxed
Government welfare based on need/approval
Society assessment based on citizen self- assessment, approval ratings, and government self-assessment

5. Problems to Solve Realistic economic cycle Feasible demographics
Identifying ramifications of different policy change

6. Simulation Optimization
Retrieve raw data and assess
Multiple variables mean the best run isn’t necessarily optimal
Optimization

7. Background Data on taxes and welfare Not much previous research
Higher taxes, more government programs
Upward trend of spending
Not much previous research
“Happiness” assessments

8. Development Q1 Q2 Q3 Q4 Preliminary research Starting the model
Finishing the model
Data handling and analysis Q3
Optimization research
Coding the optimization stage Q4
Final optimization program
Assessment of “best” policies

9. The Cycle Given Citizen traits: Wealth, wealthAssessment
Given Government traits: Wealth, wealthAssessment, approvalRating, taxRate, welfareRate, salesTaxRate
Monthly assessments to track progress

10. Approval
Government
Wealth
WealthAssessment
Approval Rating
Tax Rate
Sales Tax Rate
Responsiveness
Taxes
Welfare
Population
Wealth
Approval
WealthAssessment
Work Rate
Spending Rate
FitnessEvaluation

11. The Model
Java, JGrasp
Iterative Model allowing for multiple governments, citizen pools
Input data → Read data → Cycle → Print data → Analyze data
GNUPlot for data display
Data somewhat arbitrary now but will look for more realistic data
Optimization and randomization mitigates need for solid data

12. Why Java? Alternatives: MASON, NetLogo
Not pre-packaged, but easily modifiable
Agent based approach with outside genetic algorithm warrants relatively complex code
Handling input/output

13. Optimization Methods Genetic algorithm Hill Climbing Genetic Algorithm
Run tests
Retrieve data, determine the “best” and “breed” them
Repeat
Advantage over Hill Climbing

14. Genetic Algorithm Stochastic process Evolutionary process Problems
Crossbreed pairs with best data
Converges to local maxima/minima
Problems
Locality
Lots of variables

15. Basic Genetic Algorithm
Model
Data Output
Assessment
Sorting
Basic Genetic Algorithm
Mutation
Breeding
Selection

16. Genetic Algorithm Test
Generation 1
Generation 6

17. Specific Issues Multivariate crossover Overcoming local maxima
Tax rate (Income and sales)
Welfare rate and criteria (Responsiveness)
Overcoming local maxima
Varying degrees of importance
# of generations

18. Model Results

19. Run 1
Government Wealth
Civilian Aggregate Wealth
Economic Assessment

20. Run 2
Government Wealth
Civilian Aggregate Wealth
Economic Assessment

21. Run 9
Government Wealth
Civilian Aggregate Wealth
Economic Assessment

22. Model Discussion Dual-curve behavior – elusive
Economic assessment generally curves downwards after variation lessens
Governments gain wealth steadily
Civilians gain or lose wealth, but converge to an equilibrium

23. Model Discussion
Government wealth always increases without deceleration
Some sort of equilibrium for aggregate Citizen wealth
Assessment is erratic
Based on immediately previous data
Based on ratios, large drops/gains are “forgotten”

24. Genetic Algorithm Results
Early trial: Indication of a problem
[Trial number, generation, assessment]
0, 0,
0, 1,
0, 2,
0, 3,
0, 4, NaN
0, 5, NaN
0, 6, NaN
0, 7, NaN
0, 8, NaN
0, 9, NaN
NaN: Not a Number

25. Test runs Most result in higher assessments
Some anomalic low assessments

26. What’s going on? Ratio based assessment
Large debt or profit or zeros create calculation problems
Government basically loses too much money
Genetic algorithm doesn’t significantly result in lower assessments in most runs, but clearly there is a problem
Local maxima

27. Discussion
If the citizenry’s gains are more drastic than the government’s losses, the citizenry approves
Pattern of over-spending
People want low taxes but benefits
High spending, low revenue → debt
Government spending doesn’t necessarily help the economy
Model doesn’t change policy dynamically within a run

28. Aftermath Modifiability accomplished
Initial data not particularly positive
Many variables → data is hard to read
What is “important”?
Sustainable economies
Genetic algorithm is a success, but the model’s success ultimately lies in the assessment

29. Things to Work On A more fair assessment of the society
Current weights government and population importance equally
One group may fail but the assessment isn’t indicative if the other succeeds enough
Optimization tweaking
Test more situations – different government structures

30. Lessons Learned Variable data sets
Inferring trends and cause-effect relationships from data
A clear objective is essential