Individual-based neural-network genetic-algorithm models

apricotpigletAI and Robotics

Oct 19, 2013 (3 years and 10 months ago)

106 views

ING models:

how they work and how they are
constructed

I
ndividual based
N
eural network
G
enetic algoritm


by

Espen Strand and Geir Huse

ING models
-

Presentation layout:


Representation of individuals


Attribute and strategy vector, super
-
individual


The genetic algorithm in ING models


Structure, initiation, selection vs. variability,
reproduction


Model constraints (avoiding Darwinian monsters)


Fitness in ING models


The neural network


Network architecture, types of input, stimuli
transformation


One example of an ING model


The individuals



All individuals are numerically described by a unique strategy
vector (easy think of it as genes):



All individuals’ states are described in the attribute vector:


Strategy vector (length n)

3.2 1.3
-
0.3 2.7
-
4.1 2.3 0.1 1.0 …. n

1.6 kg 590 days 34g fat female 303 eggs

Attribute vector

Super
-
individuals


There is, depending on model complexity, an upper
practical limit to how many individuals that can be
simulated


In models where the number or biomass of individuals
are important and very high, a way around this problem
is to treat each individual as a super
-
individual


A super
-
individual simply has a number added to its
attribute vector telling how many (identical) individuals it
represents

500 ind

590 days 34g fat female 303 eggs

Attribute vector

The genetic algorithm (GA)


A GA is an algorithm that mimics evolution by natural selection.


So
-

what is required to make evolution possible?


1.
A population of individuals

2.
Genetic variability among individuals

3.
A genotype


phenotype relationship

4.
Individual variation in phenotypic success (fitness)

5.
Inheritability of genotypes from one generation to the next

6.
Introduction of new genetic variance (at least in the long run)



How is this implemented in a GA?


Implementing a GA
-

I

Ind #

Sv(1)

Sv(2)

Sv(3)

Sv(…)

Sv(n)

1

2.3

-
0.4

2.1



0.2

2

3.4

1.0

5.0



4.2

3

-
1.4

2.1

-
1.6



0.3













N

0.03

2.1

-
2.6



-
0.4

Strategy vector (length n)

Population (size N)

1

2

3

N





1.
A population of individuals

2.
Genetic variability among individuals

3.
A genotype


phenotype relationship

4.
Individual variation in fitness

5.
Inheritability of genotypes from one
generation to the next

6.
Introduction of new genetic variance

Linking behaviour to GA

1

2

Depth

Input 1

Input 2

Input 3

Input 4

Input 5

3.2 1.3
-
0.3 2.7
-
4.1 2.3 0.1 1.0
-
2.1 0.5

Behaviour

Strategy vector

Neural network

1.
A population of individuals

2.
Genetic variability among individuals

3.
A genotype


phenotype relationship

4.
Individual variation in fitness

5.
Inheritability of genotypes from one
generation to the next

6.
Introduction of new genetic variance


This link is the
cornerstone

of an
ING
-
model

Implementing a GA
-

III

500 ind

90 days

34g fat female 303 eggs

Attribute vector

0.4 ind


90 days

0.4g fat female 3 eggs

1.
A population of individuals

2.
Genetic variability among individuals

3.
A genotype


phenotype relationship

4.
Individual variation in fitness

5.
Inheritability of genotypes from one
generation to the next

6.
Introduction of new genetic variance

Implementing a GA
-

IV

+

=

or

Strategy vectors

1.
A population of individuals

2.
Genetic variability among individuals

3.
A genotype


phenotype relationship

4.
Individual variation in fitness

5.
Inheritability of genotypes from one
generation to the next

6.
Introduction of new genetic variance

About fitness (or: who gets to reproduce?)


There are two distinctly different ways to incorporate fitness in an
ING
-
model


By using a fitness measure (applied fitness)


sort all individuals in the population according to the fitness
measure and only let the fit ones reproduce. A fitness measure is
imposed on the population. Replace the old generation with the new
one. No chance of extinction. No population dynamics.


By simulating the individuals’ entire life
-
span including mortality,
gonad development, foraging, metabolic expenditure, etc…
(emergent fitness)


individuals will reproduce off
-
spring according to how well they
adapted they are to the environment. Fitness becomes an emergent
property of the model. The off
-
spring is added to the population as
juveniles and do not replace existing individuals. Emergent
population dynamics. Population may go extinct.

Model constraints


Environment






Physiology


Temperature dependent effects


Stomach limitation


Prey size limitations


Behavioural limitations


…. (this list really never ends)




GA overview

Artificial Neural Network


The basic idea of an ANN was to make an algorithm that
mimicked how a brain makes decisions based stimuli

From www.greenspine.ca/media/neuron_culture_800px.jpg


A real network of neurons

An artificial neural network (ANN)

Artificial Neural Network
-

Architecture


An ANN is constructed of:


Input


Input nodes


Input connection weights


Hidden nodes


Hidden node bias


Output connection
weights


Output node(s)

Artificial Neural Network


Input node


An input node receive a specific input and scales it
linearly to a value between 0 and 1

Artificial Neural Network


Hidden node


The hidden node sums all input connection weights
(CW) multiplied with the input node value


Artificial Neural Network


Transformation


After obtaining the value
HiddenNode
j

the value is transformed non
-
linearly. Most often a sigmoid function is used. A bias is also often
included.

HiddenNode
jT

HiddenNode
j

Artificial Neural Network


Output


The output node sums the transformed hidden node
values multiplied with the output connection weights

Artificial Neural Network


Behaviour


The value calculated by the output node(s) is used to
determine behaviour. This can be done in several ways:



Use value directly (e.g. output = swimming speed)



Use it to determine incremental step in behaviour
(e.g. NewDepth = OldDepth + output)



Transform it (sigmoid) and multiply with some
maximum range

(e.g. NewDepth = MaxDepth*output
T
)

ING
-
models: Pros and cons




Cons


No guarantee that the optimal solution is found


Need to run replicate simulations


Can be difficult to “decode” the adapted neural network


ANN = black box?


Pros


Can incorporate very high levels of complexity:


Stochasticity, Intra
-

and Inter
-
specific competition


Can be used to study emergent patterns on different levels
simultaneously:


Population dynamics, state
-
dependent behaviour


Can avoid using a measure of fitness by making fitness an
emergent property of the model.

Example: A model of a planktivours fish

Strand, E., Huse,G., Giske, J. (2002)


Time resolution


Simulates 1 day every month (and scales it to
the entire month)


Each day is divided into 5 minutes time
-
steps


Run for several hundred generations


Behaviour and life
-
history strategy


Depth position


Energy allocation


Spawning strategy


Emergent fitness


Main focus


Differences in juvenile and adult behaviour


Effects from stochastic juvenile survival on
life
-
history and behaviour

Example: A model of a planktivours fish

Vertical migration

From Baliño and Aksnes (1991)

Energy allocation

Data from Hamre
(1999)

Spawning behaviour

The End