GARP Genetic Algorithm for Rule-set Production Computational Point of View.

GARPGenetic Algorithm

for Rule-set Production

Computational

Point of View

Presentation Outline

Overview of Predictive Modeling (Arthur) GARP and DesktopGARP (Ricardo) Applications

– Climate Changes (Marinez)– Risk Assessment (Raul)– Agriculture (Victor)– Disease Systems (Town)

Geo

grap

hy

Eco

logy

Predictive Modeling Methodology

Occurrence Points

Algorithm

Precipitation

Tem

pera

ture

Ecological Niche Model

Native Predicted

Range

Predicted DistributionAfter Climate Changes

Projection Over ChangedClimate

InvasivePotential

Projection OntoAnother Region

Slilde by Town Peterson

Divide data in training data set (used to build models) and test data set (to validate the model)

Applies an algorithm to the training data set– BIOCLIM– Logistic Regression– etc.

Evaluates model quality, by asking how errors are different from random

GARP - General Approach

GARP - Data and Results

Point occurrence data

Predicted Distribution

EnvironmentalDimensions(EnvironmentalLayers)

vegetação

temperatura

precipitação

relevo

One Step at a Time

GARP = GA + RP

GA: Brief Introduction to Genetic Algorithms

RP: Rule-set Production

GA: Genetic Algorithms

Application from Artificial Intelligence

Concept taken from genetics and evolutions of species applied as a generic problem solving technique in Computer Science:

- Genes, chromossomes, mutations, insertions, deletions, crossing over, genotype, individuals, population, survival of the fittest.

For more info on GA, – visit: Marek Obitko's website at http://cs.felk.cvut.cz/~xobitko/ga/– Or ask me during the demo sessions (during lunch time)

RP: Rule-set Production

Rule: Logical Proposition Format:

If A is true then BA: preconditionB: result or prediction

Example:If Tmin_winter in [5,10] & Tavg_winter in [10,22] & Elev in [1k,2k]

then taxon is present

R: Rule-set in GARP

P: Set of training data point (spp) Elements in P: pi = (a, b) a: environmental variables at that point b: observed presence or absence

Example: p1 = (10, 12, 2k, Present)

Training Data-set

Rule-set Evaluation

P is used to test the rules in R:If a in A:

If b=B then the rule predicts correctly

If b≠B then the rule DOES NOT predicts correctly

If a not in A:

Rule does not apply to the point: test next rule

f(ri): fitness function– Percentage of points that are predicted correctly by the

rule (can be something else)

Take a Look Inside GARP

Rule Coding:r1: If Tmin_winter in [5,10] & Tavg_winter in [10,22] e Elev in [1k,2k] then presentr2: If Tmin_winter in [0,15] & Tavg_winter in [0,50] & Elev in [0,20k] then absentr3: If [Tmin_winter x 0.80 + Tavg_winter x (-0.2) + Elev x 0.45] then absent

Rule Tmin_win Tmin_win Tavg_win Tavg_win Elev Elev P/A f(r)

r1 5 10 10 22 1k 2k P 50%

r2 0 15 0 50 0k 20k A 12%

r3 0.8 --- -0.2 --- 0.45 --- A 95%

Heuristic Operators

Mutation: random modification of a gene

Before:

r2 0 15 0 50 0k 20k A 12%

r4 0 15 0 28 0k 20k A 15%

After:

Heuristic Operators

Crossing over: Exchange of segments between two chromossomes:

Before:

r1 5 10 10 22 1k 2k P 50%

r2 0 15 0 50 0k 20k A 12%

After:

r5 5 10 10 50 0k 20k P 87%

r6 0 15 0 22 1k 2k A 9%

Survival of The Fittest

Rule f(r)

r3 95%

r5 87%

r1 50%

r4 15%

r2 12%

r6 9%

Rules Sorted by f(r)

Survival of The Fittest

Rule f(r)

r3 95%

r5 87%

r1 50%

r4 15%

r2 12%

r6 9%

Threshold

Survice and have offspring

Die

Results

Rule f(r)

r3 95%

r5 87%

r1 50%

Survivors form a rule set that represents the ecological niche of that species

After <n> iterations:

Results

Ecological Niche Model of the Species:

r3: If [Tmin_winter x 0.80 + Tavg_winter x (-0.2) + Elev x 0.45] then absentr5: If Tmin_winter in [5,10] & Tavg_winter in [10,50] & Elev in [0,20k] then presentr1: If Tmin_winter in [5,10] & Tavg_winter in [10,22] & Elev in [1k,2k] then present

Rule-set projection back onto the geography space Model test, overlaying test points evaluating how

those points are predicted

Species Modeling: DesktopGARP

Acknowledgements

FAPESP and NSF BRC & NHM - The University of Kansas DesktopGARP Testers & Users Other Collaborators

DesktopGARP information on-line

Website at:

www.lifemapper.org/desktopgarp

Or Email:[email protected]

Stay With Us For More GARPing

Next: Lifemapper Project Demo Session during Lunch Time:

– Genetic Algorithms in General– GARP Algorithm– DesktopGARP live demo

In the Afternoon: Many Neat Applications

DesktopGARP

Thank you so much!!

Any questions?