PREDICTION OF IMMINENT SPECIES’ EXTINCTION

IN EcoSim

Meisam Hosseini Sedehi, Robin Gras and Md Sina

School of Computer Science, University of Windsor, Windsor, Canada

Keywords: The Species’ extinction, Demographic factor, Genetic factor, Individual-Based Model, Fuzzy Cognitive

Map, EcoSim, Feature selection.

Abstract: The process of evolution involves the emergence and disappearance of species. Many factors affect on the

survival of species. Real study of factors’ influence is particularly difficult due to the complex interaction

between them. An individual-based model (IBM) can assist in the analysis of effective factors. In this study,

using an IBM called EcoSim, we have examined the impact of some factors on the prediction of the

imminent extinction. By applying some machine learning’s techniques for feature selection and

classification, we have shown that demographic and genetic factors have a critical role for the prediction.

Especially, paying attention to both factors can highly improve the accuracy of the species’ prediction.

1 INTRODUCTION

In the course of gradual long-drawn-out evolution,

there has been an innumerable number of species;

not all of them could manage to live until now, and

some of them lived longer than the others. For a

species to survive, its individuals have to reproduce

and tolerate the environmental conditions.

The conservation of endangered species and

expansion of their longevity have always

encouraged scientists to be in search for the

fundamental reasons of species’ extinction. Each

species can be combined as one or more distinct

populations with similar ecological niche.

Populations’ extinction which is a milestone of

biology and ecology has applications in conservation

biology, biological control, epidemiology and

genetics (Drake, Shapiro, & Griffen, 2011), (Griffen

& Drake, 2008). Although much research has been

carried out in populations’ extinction and

populations’ persistence, these phenomenons are

still open questions.

There are many factors involved in extinction

that can be classified into three main realms of

Demography, Genetics and Environment (Griffen &

Drake, 2008). In real life, it is difficult to identify or

compute an exact effect of these factors separately;

it is even harder to do it altogether. Therefore,

scientists have been devising techniques, which

make it possible to compute different attributes of

species. These techniques can be separated into two

broad categories, modelling and simulation based on

the laboratory tests or the computer simulations.

Laboratory tests of extinction allow running

tests by providing simplified and tractable systems

where interactions of factors can be eliminated and

tests can be done under control. In general,

producing an apt condition and an exact repeatable

experiment is difficult, in particular, when more than

one factor has been used. Simulation techniques can

be a good alternative to inspect factors together.

In this study, we have applied a computer

simulation based on an individual-based model

(DeAngelis & Mooij, 2005) named EcoSim

(Devaurs and Gras 2010), and worked on the

prediction of species’ extinction. For this purpose,

we used different species’ attributes, linked to

demographic and genetic factors, called features.

These features were gathered from individuals who

belong to distinctive species of simulation. After

some feature selection tools have been applied, we

have used a decision tree method on the selected

features to predict if extinction occurs in a close

future. By analyzing the obtained results, we were

able to understand the effect of these features on

extinction better. We have done three experiments,

each one of them using different features set. We

have shown these features can predict species’

extinction with high accuracy.

318

Hosseini Sedehi M., Gras R. and Sina M..

PREDICTION OF IMMINENT SPECIES’ EXTINCTION IN EcoSim.

DOI: 10.5220/0003737703180323

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 318-323

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

More information at: http://sites.google.com/site/ecosimgroup

The organization of this study is as follows: In

second and third section, respectively some related

work and EcoSim are reviewed briefly. Fourth and

fifth sections contain methodologies for computation

and selection of important species’ features. Results

are discussed on sixth section. Finally, in the section

seventh, a general conclusion is given.

2 RELATED WORK

The major factors affecting on the extinction process

are Demography, Genetics and Environment

(Griffen & Drake, 2008). Demographic factors are

impacted by the population growth, the reproduction

rate and the individual’s lifespan including the

population variability, the initial population size and

the migration (Ovaskainen & Meerson, 2010).

Genetic factors correspond to a shortage of genetic

variations, which can be caused by a decrease in

fitness due to the inbreeding depression (Reed,

Lowe, Briscoe, & Frankham, 2003). Finally, factors

such as the habitat quality, the habitat fragmentation

and the environmental stressors have a major role in

extinction as the environmental factors (Patten,

Wolfe, Shochat, & Sherrod, 2007), (Drake & Lodge,

2004). The effects of most of these factors depend

on interactions with other factors and conditions

impeding the careful study of each factor.

Scientists have endeavored to clarify the impact

of extinction’s factors. For instance, Drake et al.

(2011) have experimentally investigated the effect of

population size on the extinction populations of the

flea, Daphnia Magna, in two phases: initial

population and quasi-stationarity population. They

concluded that the population size has less effect on

populations with high resilience, but the habitat size

and the environmental variability have more impact.

In another experiment, Drake and Griffen (2010)

have used laboratory populations of Daphnia Magna

to test the population dynamics due to declining

levels of food provision. They showed that the

impending extinction will be revealed by slowing

down the growth rate.

In addition to experimental works, simulation

techniques based on IBM has been used in the

simulation of ecological and evolutionary processes.

For example, Walters et al. (2002) utilized IBM to

explore the effect of demographic and

environmental stochasticity on vulnerability of Red-

cockaded Woodpecker populations. In (2007), Hovel

and Regan introduced an IBM to assess a relation

between the habitat fragmentation and the predator-

prey interactions on a group of settling blue crabs.

They have shown that the predator hunting strategy,

the prey movement and the patterns of settled prey

can modify the effect of habitat fragmentation.

Schueller and Hayes (2011) have also employed an

IBM to investigate the relationship between the

extinction risk and the minimum viable population

size for a lake sturgeon and demonstrated the

influence of inbreeding depression over the viable

population size.

IBM also can implement a speciation process

which generates new species from existing ones by

an evolutionary process. Gras, Devaurs, Wozniak, &

Aspinall, (2009) have introduced a predator-prey

ecosystem simulation called EcoSim, which

combining an IBM with Fuzzy Cognitive Map as the

behaviour model for the agents. This model allows

investigating different aspects of life by evolving

individuals in a multi-level food chain simulation.

The predators act as a pressure factor on a prey and

can be seen as an environmental stress. The prey

eats grass which availability is based on a spatial

diffusion model leading to a dynamically changing

environment. Due to the abilities of this model, we

chose it as our platform for the prediction of species’

extinction.

3 EcoSim

Gras et al. (2009) have presented EcoSim

, an IBM

including a behavioural model based on Fuzzy

Cognitive Maps (FCM) (Kosko, 1986). The

evolutionary world in this model is a grid of

1000x1000 discrete cells. Besides preys and

predators, every cell in this world may contain some

amount of grass and meat of dead prey. Predators

live on preys and preys live on grass, which is a

natural resource for preys. Every individual acts

according to its FCM which is coded in its genomes

and assigned to it at birth time. The FCM is a

directed graph containing nodes called concepts and

edges representing the influence of concepts on each

other. When a new offspring is created, it is given a

genome which is a combination of the genomes of

its parents with some possible mutations.

There are three kinds of concepts: sensitive

(such as the distance to food, to a friend and to a

sexual partner), internal (such as fear, hunger and

satisfaction) and motor (such as an escape, eat and

reproduce). In addition, each concept has an

activation level which depends on its current

perception and the past internal state. The current

activation level of a concept is computed with the

PREDICTION OF IMMINENT SPECIES' EXTINCTION IN EcoSim

319

FCM and is used to choose the next action of the of

an agent is used to compute the activation level of a

sensitive concept. The activation level of an internal

concept is influenced by the sensitive concepts of

the agent. And finally, the action of agent is selected

based on the activation levels of the motor concepts

that are affected by the sensitive and internal

concepts. Each activation level can have a positive

value between [0-1] that show an excitatory or an

inhibitory influence. EcoSim iterates continuously,

and each iteration (time step) consists in the

computation internal concepts, the choice and

application of an action for every individual. A time

step also includes the update of the world:

emergence and extinction of species and growth and

diffusion of grass. The maximum longevity of each

individual is less than 40 time steps.

This model takes advantage of a speciation

mechanism. A species is a set of individuals, which

have similar genetic characteristics (Mallet, 1995).

The species’ genome is computed by averaging the

genomes of all its members. Thus similarity can be

computed by the genetic distance between an

individual and a species. A new individual is a

member of the species of one of its parents whom

the most similar (Aspinall & Gras, 2010); unless a

hybridization event occurs, two parents are from the

same species. A speciation occurs if an individual in

a species has a genetic distance greater than a

predefined threshold from the species’ genome. As a

speciation, two new sister species are created by

splitting the species using a 2-mean clustering

algorithm. Afterward, all the individuals of the split

species are attached to one of two new species that

has fewer genetics distances. The two sister species

are initially quite similar, allowing for some

hybridization events to occur. Although they are

similar, soon they diverge genetically and became

two completely independent species. The

hybridization happens in the nature, although the

effective mechanisms on and long-time impact of it

is unclear (Gilman & Behm, 2011).

4 COMPUTATION OF

ATTRIBUTES

For predicting the species’ extinction based on

demography and genetics factors, we need to find

features of all species. In EcoSim, every individual

has a certain number of attributes, which are fully

quantiﬁable at each time step of the simulation.

Moreover, each species’ attribute is the average

value of this attribute for all individuals belonging to

that species. Let ={



,



, ..., 



} be the set

ofm species at a given time step, where each 



is a

set of individuals ={i





,...,i



}. The value of m

may vary over time since species can emerge or

extinct. The same is true for p as individuals can be

born or die at any time step. In addition,

assume



={a





, ..., a



} is the set of n

attributes of i



individual. We define AS



(

)

as a

function that maps one attribute of individuals to one

attribute of species by equation,



(k) =



a



∀







(1)

In addition to these features, a species has few

specific attributes that do not apply to a single

individual, such as the spatial diversity, the genetic

diversity and the number of species per step time.

The spatial diversity measures the distribution of

individuals in one species based on the locations of

all its individuals and is computed in two steps:

computation of a spatial centre of the species and the

spatial standard deviation. EcoSim’s world is a torus

which means that the opposite boarders of the grid

are adjacent. Therefore, we apply the circular

statistics (Jammalamadaka & Sengupta, 2001) to

compute the centre of the species’ spatial

distribution.

The genetic diversity of a species measures how

much diversity exists in the gene pool of the

individuals of a species. The entropy measure is

commonly used as an index of diversity in ecology

and increasingly used in genetics (Sherwin, 2010).

As a result, the genetic diversity is the entropy of the

genomes of all individuals belonging to a species,

and represents the level of similarity between all the

genomes of a species.

5 FEATURE SELECTION AND

DATA ACQUISITION

Preparing one dataset of features takes as long as

one month because we have to run a complex

simulation, in which the number of individuals is

often more than two hundred thousand at a time.

EcoSim generates massive raw data about the

individuals and the species in each time step. In this

study, we have focused on prediction of prey’s

species extinction. We have created a dataset, each

sample of the dataset shows the information about

one species at a given step time. Moreover, for

making a prediction, each sample is labelled positive

or negative respectively, meaning that extinction

ICAART 2012 - International Conference on Agents and Artificial Intelligence

320

will appear during the next 50 time steps or not.

The raw dataset includes 52 features, combining

demographic and genetic factors. These features of

each species contain: 12 sensitive concepts’ average

activation level, 7 internal concepts’ average

activation level, 7 motor concepts’ average

activation level, 11 actions frequency, the population

size, the ratio of individuals to whole population, the

number of dead individuals, the genetic diversity,

the spatial diversity, the average of age of

individuals, the energy and speed of individuals, the

average of failed reproduction events based on age,

the genetic distance and energy of individuals, the

average of genomes of individuals, the average

genetic distance from species’ genome and the

average amount of energy transmit from a parent to

a child. There is also a global feature representing

the current number of species.

Table 1: List of the 22 selected features for each species

contains three parts: 1) demographic factors, 2)

individuals’ perceptions and actions 3) the genetic, energy

and age factors.

Part # Features

Number of species

Population size

The ratio of individuals to whole population

Spatial diversity

Number of dead individuals

Perception of predators’ distance (sense.)

Perception of food’ distances (sense.)

Perception of friend’ distances (sense.)

Perception of amount of food (sense.)

Perception of amount of mating’s partner

(sense.)

Rate of prey’ escape from predators (act.)

Rate of search for food (act.)

Rate of exploration inside the world (act.)

Rate of eating food (act.)

Rate of reproduction (act.) [percentage of

individuals breeding]

Rate of failed reproduction (failed act.)

Perception of amount of energy (sense.)

Average of age

Average energy

Average of age failed reproductions events

Average energy reproductions events

Genetic diversity

These amounts of features led to slow

classification with the lower level of accuracy.

Moreover, we tried to work on the features having

more impact on classification. Therefore, we have

reduced the number of features with machine

learning’s techniques. For this purpose, different

methods such as simple Genetic Search (crossover-

prob: 0.6, mutation-prob: 0.03, population-size: 20)

(Goldberg, 1989), Best First Searches (the space of

attribute subsets by hill-climbing augmented), and

Greedy (stepwise) Search have been tested on

WEKA (V3.6.4). A weighted linear combination of

their results leads to the selection of 22 features that

will be used for learning the prediction model. The

list of these selected features has been shown in

Table 1.

In our dataset, most of the samples are labelled

negative, because in most of the time steps a given

species will not become extent in the next 50 time

steps that makes a bias in the number of negative

samples to positive samples. Then, we divided the

initial dataset into a training set made of 80 percents

of the positive samples plus two times the same

amount of randomly selected negative samples and

the rest of the samples as the testing set.

6 EXPERIMENTAL RESULTS

The prepared dataset contains 120,000 samples

related to about 150 species coming from 10000

time steps of one unique run. The training set has

formed about 24,000 samples contains 8000 positive

samples and 16000 negative samples. For comparing

the quality of classification based on different

experiences, four measures of accuracy, true positive

(TP) rate, true negative (TN) rate, global accuracy,

and ROC area have been used. The global accuracy

shows the percentage of correctly classified samples.

The true positive (negative) rate presents the

percentage of true classified positive (negative)

samples. Finally, ROC area reveals sensitivity by

measuring the fraction of true positives out of the

positives versus the fraction of false positives out of

the negatives.

We chose to use decision tree with the

confidence factor 0.25 for pruning and 100

minimum instance per leaf (Quinlan, 1993) as a tool

for prediction, because it reduced the number of

produced rules and therefore, allows an analysis of

the produced model. This reduction in the number of

rules has only decreased very slightly the prediction

accuracy but has increased a lot the interpretability

of the rules. However, we also tested several other

learning techniques, and they all led to similar

results. We have arranged three experiments for

training a decision tree based on three parts of

selected features of Table 1. These three parts

contain respectively demographic factors, features

related to individuals’ perception and action, and

finally a combination genetic factor and features

linked to energy and age.

To test how general the predictors we have

obtained, and to validate our results, we have

PREDICTION OF IMMINENT SPECIES' EXTINCTION IN EcoSim

321

Table 2: Results of train set, test set and validation set.

Experience Dataset Accuracy TN Rate FN Rate TP Rate FP Rate ROC Area

First

Train 80.16% 0.76 0.11 0.88 0.23 0.84

Test 75.11% 0.74 0.12 0.87 0.25 0.83

Validation 67.78% 0.65 0.11 0.88 0.34 0.77

Second

Train 92.02% 0.95 0.15 0.84 0.05 0.96

Test 94.90% 0.95 0.15 0.84 0.04 0.96

Validation 90.30% 0.91 0.23 0.76 0.08 0.92

Third

Train 92.97% 0.95 0.13 0.87 0.04 0.95

Test 95.33% 0.95 0. 14 0.85 0.04 0.95

Validation 91.81% 0.92 0.17 0.82 0.09 0.92

performed an independent run of the simulation, to

generate a new dataset. This dataset consists of

about 30,000 positive samples and 280,000 negative

samples.

In the first experiment, we trained a decision

tree using only features that can be provided by

observation in the real ecosystem to see how good

they are for prediction alone. This set is made up of

five features: the number of species, the population

size, the ratio of individuals to whole population, the

spatial diversity and the number of dead individuals

that present demographic factors.

It can be seen in Table 2 that, with these five

features, the accuracy of the species’ prediction is

about 80%. The high true positive rate implies these

features can alert for the possibility of extinction.

Furthermore, the test set confirms the validity of

selected features in prediction of survival or

extinction. However, results of the validation set are

not as good as test set; and they imply that these

features are not sufficient as a general predictor.

According to the decision tree’s results (not

shown here), spatial diversity has a great impact on

accuracy. Based on ecological studies, this

phenomenon was expectable (Griffen & Drake,

2008), although population size did not have a

considerable effect. This weak influence on

prediction quality can be due to a quasi-stationarity

population (Drake et al., 2011).

In the second experiment, in addition to the

features of the first experiment, the second set of

features has been used. These 11 features include the

rate of actions of individual and their sensitive

concepts. The results of the second experiment have

been shown in Table 2. The second features have

increased the prediction of species’ extinction from

80% accuracy to more than 90% accuracy. The

value of true positive rate and ROC area reveals that

the second set of features are highly associated with

an occurrence of extinction. The high accuracy on

the test set and validation set prove the validity of

this experiment. In this experiment, in addition to

the spatial diversity, features: the rate of failed

reproduction, the perception of predators’ distance,

the perception of the amount of food, the perception

of the amount of mating’s partner, the rate of eating

food and the rate of prey’ escape, are the ones that

have the more important effect on prediction

accuracy.

In the third experiment, all selected features have

been applied in prediction. The added features

mostly are a representation of the characteristics and

genetic traits of an individual. Based on Table 2,

using all features together leads to a better prediction

than what is obtained in first and second

experiences. Especially, value of accuracy, true

positive rate and true negative rate all show an

increase. More interestingly, the increase accuracy

in validation set demonstrates that these gains are

not due to over fitting problem. Generally, validation

set reveals that these features correspond to general

characteristics, which can prognosticate species’

trend toward extinction in quite diverse situations.

As a result, the more accurate determination of

extinction is obtainable through the combination of

various features which means that the association of

demographic and genetic factors can increase the

power of diagnosis an imminent extinction.

According to the decision tree, genetic diversity, the

average age of failed reproductions events and

average of age have the more impact on extinction’s

prediction.

7 CONCLUSIONS AND FUTURE

WORK

Species’ extinction is affected by demographic,

genetic and environmental factors. Analysis of

effects of these factors in real life due to the

complex interaction between these factors is very

hard. IBM models are an alternative way for scrutiny

the effects of different factors.

In this study, we used EcoSim, an IBM with

genetic traits, for finding critical features of

ICAART 2012 - International Conference on Agents and Artificial Intelligence

322

extinction’s prediction. This model allows us to

work on numerous factors simultaneously. Based on

this model we set up three experiences with different

species’ features on two datasets. Results confirmed

the impact of demographic and genetic factors on

prediction of species extinction and showed that

very good predictor can be built. We demonstrated

that a combination of these factors can improve the

prediction’s accuracy. Moreover, the accuracy of

validation set presented the general ability of

selected features in prediction of impendent

extinction of species.

In a next step, we want to focus on correlation

and dependency between features. For this purpose,

we have to work on the analysis of features’

interactions and on the extraction of biologically

significant rules. These rules will help to reveal the

priority and relation between features and provide

some insight about the biological mechanisms

involved in species’ extinction.

ACKNOWLEDGEMENTS

This work is supported by the NSERC grant

ORGPIN 341854, the CRC grant 950-2-3617 and

the CFI grant 203617 and is made possible by the

facilities of the Shared Hierarchical Academic

Research Computing Network (SHARCNET, www.

sharcnet.ca).

REFERENCES

Aspinall, A., & Gras, R. (2010). K-Means Clustering as a

Speciation Mechanism within an Individual-Based

Evolving Predator-Prey Ecosystem Simulation. Active

Media Technology, LNCS6335, 318-329.

DeAngelis, D. L., & Mooij, W. M. (2005). Individual-

Based Modeling of Ecological and Evolutionary

Processes 1. Annual Review of Ecology, Evolution,

and Systematics, 36(1), 147-168.

Devaurs, D., & Gras, R. (2010). Species abundance

patterns in an ecosystem simulation studied through

Fisher’s logseries. Simulation Modelling Practice and

Theory, 18(1), 100-123. Elsevier B.V.

Drake, J. M., & Griffen, B. D. (2010). Early warning

signals of extinction in deteriorating environments.

Nature, 467(7314), 456-9. Nature Publishing Group.

Drake, J. M., & Lodge, D. M. (2004). Effects of

environmental variation on extinction and

establishment. Ecology Letters, 7(1), 26-30.

Drake, J. M., Shapiro, J., & Griffen, B. D. (2011).

Experimental demonstration of a two-phase population

extinction hazard. Journal of the Royal Society,

Interface / the Royal Society, 8(63), 1472-9.

Gilman, R. T., & Behm, J. E. (2011). Hybridization,

Species Collapse, and Species Reemergence After

Disturbance To Premating Mechanisms of

Reproductive Isolation. Evolution, no-no.

Goldberg, D. E. (1989). Genetic algorithms in search,

optimization, and machine learning. Addison-Wesley

Professional.

Gras, R., Devaurs, D., Wozniak, A., & Aspinall, A.

(2009). An individual-based evolving predator-prey

ecosystem simulation using a fuzzy cognitive map as

the behavior model. Artificial life, 15(4), 423-63.

Griffen, B. D., & Drake, J. M. (2008). A review of

extinction in experimental populations. The Journal of

animal ecology, 77(6), 1274-87.

Hovel, K. a, & Regan, H. M. (2007). Using an individual-

based model to examine the roles of habitat

fragmentation and behavior on predator–prey

relationships in seagrass landscapes. Landscape

Ecology, 23(Sep1), 75-89.

Jammalamadaka, S. R., & Sengupta, A. (2001). Topics in

circular statistics (Vol. 5). World Scientific Pub.

Kosko, B. (1986). Fuzzy cognitive maps. International

Journal of Man-Machine Studies, 24(1), 65-75.

Elsevier.

Mallet, J. (1995). A species definition for the modern

synthesis. Trends in Ecology & Evolution, 10(7), 294-

299. Elsevier.

Ovaskainen, O., & Meerson, B. (2010). Stochastic models

of population extinction. Trends in ecology &

evolution, 25(11), 643-652. Elsevier Ltd.

Patten, M. A., Wolfe, D. H., Shochat, E., & Sherrod, S. K.

(2007). Habitat fragmentation, rapid evolution and

population persistence. Evolutionary Ecology, 7, 235-

249.

Quinlan, J. R. (1993). C4. 5: programs for machine

learning. Morgan Kaufmann.

Reed, D. H., Lowe, E. H., Briscoe, D. A., & Frankham, R.

(2003). Inbreeding and extinction: Effects of rate of

inbreeding. Conservation Genetics, 4(3), 405-410.

Schueller, A. M., & Hayes, D. B. (2011). Minimum viable

population size for lake sturgeon (Acipenser

fulvescens) using an individual-based model of

demographics and genetics. Canadian Journal of

Fisheries and Aquatic Sciences, 68(1), 62-73.

Sherwin, W. B. (2010). Entropy and Information

Approaches to Genetic Diversity and its Expression:

Genomic Geography. Entropy, 12(7), 1765-1798.

Walters, J. R., Crowder, L. B., & Priddy, J. A. (2002).

Population Viability Analysis for Red-Cockaded

Woodpeckers Using an Individual-Based Model.

Ecological Applications, 12(1), 249-260.

WEKA, V3.6.4, http://www.cs.waikato.ac.nz/ml/weka/

PREDICTION OF IMMINENT SPECIES' EXTINCTION IN EcoSim

323