Forrester System Dynamics

based on observation + physical insight

semi-physical, semi-inductive methodology

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

1/20

Methodology
1. levels/stocks and rates/ows

dP dt
Level population inventory money

BR

DR
Outow death rate sales expenses

Inow birth rate shipments income

2. laundry list: levels, rates, and inuences

birth rate

birth

population

3. Inuence Diagram (+ and -) 4. Structure Diagram

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

2/20

Causal Relationships

graduates

standard of living (SOL)

SOL

graduates

beer consumption time beer consumption

latent variable

SOL

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

3/20

Archetypes

Bellinger http://www.outsights.com/systems/

structure diagrams Common combinations of reinforcing and balancing structures

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

4/20

Archetypes: Reinforcing Loop

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

5/20

Archetypes: Balancing Loop

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

6/20

Archetypes: Escalation

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

7/20

Archetypes: Escalation Unrolled

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

8/20

Archetypes: Accidental Adversaries

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

9/20

Archetypes: Tragedy of the Commons

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

10/20

Forrester System Dynamics

uptake_predator loss_prey Grazing_efficiency

prey_surplus_BR

predator_surplus_DR

Predator

Prey

2species predatorprey system

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

11/20

Inductive Modelling: structure

BR = f (POP, POL, MSL, ...) BR = BRN.f1(POP, POL, MSL, ...) BR = BRN.POP.f2(POL, MSL, ...) BR = BRN.POP.f3(POL).f4(MSL). ... f3(POL) = "inversely proportional" f4(MSL) = "proportional"

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

12/20

Structure Characterisation: LSQ t

X

2 X(t) = - gt /2 + v_0 t X(t) = A sin (b t )

t 2 LSQ (sin) < LSQ (t )

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

13/20

Feature Extraction

1. Measurement data and model candidates 2. Structure selection and validation 3. Parameter estimation 4. Model use

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

14/20

Feature Rationale

Minimum Sensitivity to Noise Maximum Discriminating Power

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

15/20

Throwing Stones

Candidate Models

1. x 2. x

1 2 2 gt

v0t

Asin bt

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

16/20

Feature 1 (quadratic model)

McGill, 29 November, 2000

hv@cs.mcgill.ca

Feature 2 (sin model)

solve numerically for b

F2

200

McGill, 29 November, 2000

hv@cs.mcgill.ca

bA bA

bB bB

CS 308-767A Modelling and Simulation

1 tg bt b

F1

gA gB

CS 308-767A Modelling and Simulation

xi xi

gi

2xi ti2

2xi i ti

AB

17/20

18/20

Feature Space Classication

feature sin

F2 = 200 |bA -bB|/(bA + bB) 1/b(tg(bt)) = xi/xi_der

2 Feature t

F1 = gA/gB gi = 2xi/ti^2 - 2xi_der/ti

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

19/20

Pandemomium Model

edge

vector templates

CCD

vectorise

stroke tracing

threshold smooth thin background neural net strokes convex deficiences voting scheme

pixels

"image" demon

"feature" demon

"cognitive" demon

"decision" demon

McGill, 29 November, 2000

hv@cs.mcgill.ca

CS 308-767A Modelling and Simulation

20/20