Download Genetic Algorithms: Genetic Algorithm Applications to Actuarial Problems ARC 2012

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Vectors in gene therapy wikipedia , lookup

Polymorphism (biology) wikipedia , lookup

Gene wikipedia , lookup

Genetically modified food wikipedia , lookup

Gene therapy wikipedia , lookup

Genome evolution wikipedia , lookup

Twin study wikipedia , lookup

Artificial gene synthesis wikipedia , lookup

Koinophilia wikipedia , lookup

Quantitative trait locus wikipedia , lookup

Site-specific recombinase technology wikipedia , lookup

Biology and consumer behaviour wikipedia , lookup

Genetic code wikipedia , lookup

Pharmacogenomics wikipedia , lookup

Genetic drift wikipedia , lookup

Medical genetics wikipedia , lookup

Behavioural genetics wikipedia , lookup

Heritability of IQ wikipedia , lookup

Gene expression programming wikipedia , lookup

Designer baby wikipedia , lookup

Population genetics wikipedia , lookup

Human genetic variation wikipedia , lookup

History of genetic engineering wikipedia , lookup

Genetic testing wikipedia , lookup

Genetic engineering wikipedia , lookup

Public health genomics wikipedia , lookup

Genetic engineering in science fiction wikipedia , lookup

Microevolution wikipedia , lookup

Genome (book) wikipedia , lookup

Transcript
ARC
2012
Genetic Algorithms: Genetic Algorithm
Applications to Actuarial Problems
Dave Snell, ASA, MAAA, CLU, ChFC, FLMI, MCP, ARA, ACS
Technology Evangelist, Vice Chairman’s Office
RGA Reinsurance Company
Genetic Algorithms – 4-AUG-2012 Dave Snell
SOCIETY OF ACTUARIES
Why do we call these Genetic Algorithms?
They mimic our current knowledge of
genetics.
We have trillions of cells.
DNA represents a blueprint for a cell.
It is used to generate copies.
The actual process involves proteins and
lots of other biological terms …
Genetic Algorithms – 4-AUG-2012 Dave Snell
Mitosis – exact copies
Meiosis – copy portions from two parents
Mutation – inexact copy – rarely lives
Over time - a long time - we evolve
… taller, less hairy, and more adept at making tools.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Very basic genetics
§  Humans have about 25,000 genes made from A-T and C-G pairs
§  DNA strand – double helix of two strands – each about 1.8 meters
§  In meiosis, the strand separates into 46 chromosomes (in 23 pairs)
§  Alleles (forms of a gene) help determine physical or behavioral traits
Genetic Algorithms – 4-AUG-2012 Dave Snell
Does the size of a Genome determine species dominance?
NO! – It’s all in how you use it!
§  Human 3,200,000,000 (3.2Gb)
§  Lungfish130,000,000,000 (130Gb)
§  "Amoeba" 670,000,000,000 (670Gb)
Genetic Algorithms – 4-AUG-2012 Dave Snell
Where is the Research being done on Genomes?
All over the world; but especially …
National Institutes of Health (NIH)
“This
is why I go to work!”
RMS – rhabdomyosarcoma
-Frederick S. Huang, MD
The Genome Institute at
Washington University
Genetic Algorithms – 4-AUG-2012 Dave Snell
Genetic Algorithms
Meet Amoeba Robby
Robby is only able to do 7 actions:
0-Move North 1 square
1-Move South 1 square
2-Move West1 square
3-Move East 1 square
4-Stay put
5-Pick up a can
6-Make random move (from 0-5)
Our first Robby is blind!
Story of Robby simplified from Complexity: A Guided Tour by Melanie Mitchell, Ph.D.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Genetic Algorithms (cont.)
§  Robby actions can be represented by a strategy
0- north 1
string, like ‘35265’
1- south 1
(move East, pickup, move South, random, pickup) 2- west 1
3- east 1
4-Stay put
5-Pick up
6-random
§  In 100 steps, what strategy would you use to pick
up as many randomly placed cans as possible?
You get 10 points for each can.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Genetic Algorithms (cont.)
§  Robby actions can be represented by a strategy
0- north 1
string, like ‘35265’
1- south 1
(move East, pickup, move South, random, pickup) 2- west 1
§  In 100 steps, what strategy would you use to pick
up as many randomly placed cans as possible?
You get 10 points for each can.
3- east 1
4-Stay put
5-Pick up
6-random
• Oh yeah! … you lose 5 points every time
you hit a wall.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Try all possibilities?
7^100 > 10^84 so trying all
possibilities might take too long
… compare to
around 10^17 seconds since
Big Bang billions of years ago
Genetic Algorithms – 4-AUG-2012 Dave Snell
Hill Climbing … works really well when there is only one hill
Mount Everest?
Our family in Tibet
Yak, yak, yak
Genetic Algorithms – 4-AUG-2012 Dave Snell
Genetic Algorithms (a simplistic recipe)
1.  Randomly assign
the gene string for
100 robots.
2.  Run 100 tests with cans scattered randomly.
3.  Score each robot and eliminate losers
(survival of the fittest).
4.  Assign mating privileges to survivors per
test scores.
In a kangaroo mob, only
the highest ranking male
gets mating privileges
Genetic Algorithms – 4-AUG-2012 Dave Snell
Genetic Algorithms (simplistic recipe continued.)
5.  Pick parents, and pair them off to produce the
next generation
Parent M (52.80 points): 54335351253404 315142153601520652551511513145155663
Parent F (45.60 points) :54335351525534 633242153604216362551511503145155625
arbitrary (pseudo-random) split point
Child A: 54335351253404 633242153604216362551511503145155625
genes from parent M
genes from parent F
Child B: 54335351525534 315142153601520652551511513145155663
Genetic Algorithms – 4-AUG-2012 Dave Snell
Genetic Algorithms (simplistic recipe wrap-up.)
1. 
2. 
3. 
4. 
5. 
6. 
7. 
Randomly assign the initial gene strings.
Run tests.
Score each robot and eliminate losers.
Assign mating privileges to highest scorers.
Pick parents, and produce next generation.
Randomly mutate a few genes (actions).
Repeat for 100 (or 1,000 or 10,000) generations.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Demonstration: Blind Robby
Watch evolution in action
Genetic Algorithms – 4-AUG-2012 Dave Snell
Genetic Algorithms
(How twins are made)
Genetic Algorithms – 4-AUG-2012 Dave Snell
Let’s give Robby a sense of Vision
Robby can see five squares:
the current square.
the square immediately to the East
the square immediately to the South
the square immediately to the West
the square immediately to the North
Assume each of these five squares can have three different states:
0 = Empty; 1= Contains a can; 2 = Wall
The total number of visual cases is
(3 states/cell) ^(5 visible cells) = 243
Genetic Algorithms – 4-AUG-2012 Dave Snell
Robby with Vision (continued)
Let’s number these 243 cases from 0 to 242
3^4
North
3^3
West
3^2
South
3^1
East
3^0
Current
Case
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
1
2
0
0
1
2
3
2
2
2
2
2
2
2
2
1
2
241
242
…
Case
Robby is in a wall
(not possible)
other impossible cases
Action
0 MoveNorth
1 PickUpCan
2 MoveSouth
3 MoveNorth
241 MoveEast
242 MoveWest
It doesn’t matter that some of the cases are not physically possible.
Our case table mathematics is not limited by the possible!
Generation 1 will have a randomly generated column of actions.
Future generations inherit actions from parents.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Robby with Vision (continued)
‘
Suppose we want the action for this situation:
Robby sees a wall to the
East and cans to the
North and South.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Robby with Vision (continued)
This would correspond to case number 96 in the virtual case table,
calculated as follows:
0*3^0 (current) + 2*3^1 (East) + 1*3^2 (South)
+ 0*3^3 (West) + 1*3^4 (North) = 0+6+9+0+81 = 96
CASE TABLE
Case Number Current
East
0
0
1
0
…
96
0
…
242
2
South
West
North
0
0
0
0
0
0
0
1
2
1
0
1
2
2
2
2
Genetic Algorithms – 4-AUG-2012 Dave Snell
Action
MoveEast
MoveEast
…
MoveNorth
…
PickUpCan
For this particular ‘gene string’, case
number 96 results in action
MoveNorth. Note: The actions for any
particular gene string do not have to
make sense! Only the better ones
survive and propagate.
Demonstration: Robby with vision
Watch evolution in action
Genetic Algorithms – 4-AUG-2012 Dave Snell
Mitosis & Meiosis
Our Elites correspond to
Mitosis
Genetic Algorithms – 4-AUG-2012 Dave Snell
The rest of the new generation
are analogous to Meiosis
Meiosis in action - Genetics 0.001
Parent 'M'
Gene
Parent 'F'
Child1
Child2
Child3
Child4
1
Dad
Mom
Dad
Mom
Dad's Dad Mom's Dad
Dad's Mom Mom's Mom
Dad's Dad Mom's Mom
Dad's Mom Mom's Dad
2
Dad
Mom
Dad
Mom
Dad's Mom Mom's Mom
Dad's Mom Mom's Dad
Dad's Mom Mom's Mom
Dad's Mom Mom's Dad
3
Dad
Mom
Dad
Mom
Dad's Dad Mom's Mom
Dad's Mom Mom's Mom
Dad's Mom Mom's Mom
Dad's Mom Mom's Dad
4
Dad
Mom
Dad
Mom
Dad's Dad Mom's Dad
Dad's Dad Mom's Mom
Dad's Mom Mom's Mom
Dad's Mom Mom's Mom
…
5,000
5001
…
Note that each child gene is from two of the four
grandparents: one on Mom’s side, one on Dad’s side
(excluding mutations)
Dad
Mom
Dad
Mom
Dad's Mom Mom's Mom
Dad's Dad Mom's Mom
Dad's Dad Mom's Mom
Dad's Mom Mom's Mom
Dad
Mom
Dad
Mom
Dad's Mom Mom's Dad
Dad's Mom Mom's Dad
Dad's Mom Mom's Dad
Dad's Dad Mom's Dad
set of genes from
Parent genes prior to meiosis – eachDad
person
gets one set from Dad and another set from
Mom
Genetic Algorithms – 4-AUG-2012 Dave Snell
set of genes from
Mom
Genetic Algorithms are not limited to two parents
(or even four grandparents)
Your ‘genes’ can have
vastly different
characteristics and
components.
A ‘child’ can have
genes drawn from
several different
parents.
Genetic Algorithms – 4-AUG-2012 Dave Snell
That was fun! … but let’s see a Health Insurance
application of Genetic Algorithms.
Brian Grossmiller’s application:
How to find a more cost effective set of health
providers and still have adequate coverage.
Genetic Algorithms – 4-AUG-2012 Dave Snell
Demonstration: Health Provider Network
Gene$c Algorithms – 4-­‐AUG-­‐2012 Dave Snell But does it work?
§  Overall Network Relativity: 1.00
§  Actuaries (using standard optimization
techniques): 0.72
§  Genetic Algorithm: 0.54
(lower number is better … a lot lower is a lot better)
Genetic Algorithms – 4-AUG-2012 Dave Snell
Basics to take home with you
§  Keep problem manageable
§  Define objective
§  Robby - clean world
§  Provider Network - lower health costs and adequate coverage
§  Define universe for the actors
§  Robby - 10x10 grid littered with soda cans
§  Provider Network - all available providers, their specialties, their relative costs
§  Define actions – how actors change the universe
§  Blind Robby – deterministic gene string with 7 possible actions for each gene
§  Robby with vision – adapts to environment and chooses actions (from 7)
§  Provider network – deterministic gene string, simple IN or OUT of network
§  Create evaluation/ranking scheme (points for cans, or relative cost)
§  Make a reproductive model
§  Blind Robby and Robby with Vision – 2 parents and a crossover point
§  Provider network - N parents and pseudo-random selection gene-by-gene
§  Validate (save entire gene string for best set and verify it works)
Genetic Algorithms – 4-AUG-2012 Dave Snell
Wrap-up, and Feedback
Most actuarial presentations are like mathematics, 1/2 the people can’t
understand them …
and the other 2/3 don’t care about them
Good tools and
Fun? …
Bonus video 1 - Robby
… or did we waste your time?
Bonus video 2- EC gain
Bonus video 3 – EC strategy
Let’s Talk!
Genetic Algorithms – 4-AUG-2012 Dave Snell
SOCIETY OF ACTUARIES
ARC
2012
Genetic Algorithms: Genetic Algorithm
Applications to Actuarial Problems
Dave Snell, ASA, MAAA, CLU, ChFC, FLMI, MCP, ARA, ACS
Technology Evangelist, Vice Chairman’s Office
RGA Reinsurance Company
Genetic Algorithms – 4-AUG-2012 Dave Snell
SOCIETY OF ACTUARIES