You are on page 1of 8

Genetic algorithms: how can a machine optimize a

set of parameter
21/09/21

Ex: generate parameters to choose the best configuration for software automatons

Solve the optimization of sintering conditions in a solution space of dimension 17;

Solve an “ill posed inverse problem” about reconstruction of volume from 2D pictures and we’ll see
that in more details as an example in a few minutes

eeeeee
eeeeee

In real life
Mitosis
Ie DNA -> DNA’ (clone of himself)
 No a lot of diversity
 Efficient if you have to be specific in a specific situation

Sexual reproduction
DNA1+DNA2 = DNA3
 A lot a diversity : constantly mixed source

 DNA is information
 Replication mean copy, copy mean errors (corrected or not)
 Copying and mixing create new information, more or less adapted
 A population of individuals is a lot of different info, and info diversity is valuable
 Why not consider a “population” of “genes/individual” (that are in fact information) that
could be “improved” w/ time by mimicking natural behaviours ?
MODELISATION 3D DE LA MICROSTRUCTURE

going from 3D to 2D

3D -> 2D OK

The other ways iis a pb bc n

Optimization algorithm

Schema

Try to modify original 3D

Use of genetic algotithm

(information)
0. Individual need to have values
Rather good/bad solution
1. Selection (doesn’ automatically select the best individual bc it on probability, this is imortante to
protect diversity)
2. Crossover between the parent (individual selected)
3. Mutation
4. Reinjection

>>> stability of the convergence


>>> multiple solutions are not a problem
Sampling the solution space

VALIdation of the genetic algotithm

Robust algo, diff object shapes are not a pb, analalytic or simulated model
GENETIC ALGORITHMS

Inspired by natural adaptation behaviours


Based on information itself, nothing else needed
No direct link w/ the pb itself
Need to evaluate the quality of a result/an individual
Able to generate multiple solutions at the same time

Random, …

Set of possible select two of crossover/


mutation keep or reject
solution these solution

One or more
solutions

>>> Stop or keep evolving


28/09/21

Random, …

Set of possible
solution

We can’t afford to explore the all the possibilities


Try to explore solution by choosing solution at random (but uniformely)
(>god thing to preserve bio diversity)

1 phenotype = 1 solution

select two of
these solution

Being able to value each solution => fitness function

“if you cannot measure it, you cannot improve it”

Probabilistic selection mandatory to preserve genetic diversity / information diversity.

 Don’t want to specialise to much (at least at the beginning)

crossover/
mutation

Mix information (crossover)

Sometimes alter some information arbitrarily (mutation)


keep or reject

Value the new solution => fitness function

Chose how to (re)place information in the global set.

 Improve the best case


 // the average case
 Allow multiple aggregation points (make appear multiple success / solution)
GENTEIC ALGORITHMS : IMPLEMENTATION DETAILS

Fitness function

For each inividual : possibility to

[information individual] = > fitness value

Is this combinaison good or not ? FV dit ca

FV function de chaque individu

(Perhaps not selective too much)

Not too brutal (at least at the beginning) : do not make diseappear info to soon (preserve diversity)

But at the same time : had to be selective enough

Finding a good metric might be difficult

 Sometimes, change the fitness function after a while

Give scalar value to individue

How an we choose efficiently our 2 parents ?

Probabilistic selection

Each individual can be selected, but there a (powerful) bias

|1| 2 | 3 | 4 |5| 6 |
3, 6 : strong probability of selection

1, 5 : low probability of selection

Fitness value

| 57 | 140 | 278 | 93 | 28 | 206 |

Cumulated values

| 57 | 197 | 475 | 568 | 596 | 802 |


Random value : 691 ↑
Crossover

Parent : | B | B | B | B | B | B | B | & |O|O|O|O|O|O|O|

Mixing / masking : | B | B | B | B | O | O | O |

Complex mix / mask : | B | B | O | B | O | O | O | (at random)

 Create new combinaison

Linear interpolation : | G1 | G2 | ..

 Create nex valeus

Mutation

Mixing is not enought

 Binary reversal (bit flip)


 Inject a value
 Add / substract ( a value)

|B|B|G|B|O|O|O|

 Not every time


 Too much randomness => random research (generate individual a random)

Reinject

Replace the worst individual in a population (good for convergence)

Replace at random : easier to doo (computer pov)

Add it to the global set /or/ create a new population


GENTEIC ALGORITHMS : PRATICAL EXEMPLE

[GA.example]
Things to take car of

 The devils is in the details


 Why were some choices made like this
 Identify common traps
 Going back to the power […??]
 [??]

Let’s add some environ contraint

You might also like