Ants in the Pants!

An Overview
 Real world insect examples
 Theory of Swarm Intelligence
 From Insects to Realistic
A.I. Algorithms
 Examples of AI applications
Bees
 Colony cooperation

 Regulate hive temperature

 Efficiency via
Specialization: division of
labour in the colony

 Communication : Food
sources are exploited
according to quality and
distance from the hive
Termites
 Cone-shaped outer
walls and
ventilation ducts
 Brood chambers in
central hive
 Spiral cooling
vents
 Support pillars
Ants
 Organizing highways to and from
their foraging sites by leaving
pheromone trails

 Form chains from their own bodies

to create a bridge to pull and hold
leafs together with silk

 Division of labour between major

and minor ants
Social Insects
 Problem solving benefits include:
 Flexible
 Robust
 Decentralized
 Self-Organized
Summary of Insects
 The complexity and sophistication of
Self-Organization is carried out with no clear leader

 What we learn about social insects can be applied to the

field of Intelligent System Design

 The modeling of social insects by means of

Self-Organization can help design artificial distributed
problem solving devices. This is also known as Swarm
Intelligent Systems.
Interrupt The Flow
The Path Thickens!
The New Shortest Path
Adapting to Environment
Changes
Four Ingredients of
Self Organization

 Positive Feedback
 Negative Feedback
 Amplification of Fluctuations -
randomness
 Reliance on multiple interactions
Types of Interactions
For Social Insects
 Direct Interactions
 Food/liquid exchange, visual contact,
chemical contact (pheromones)

 Indirect Interactions (Stigmergy)

 Individual behavior modifies the environment,
which in turn modifies the behavior of other
individuals
WEB CLUSTERING
Why?
The size of the internet has doubling its size every year.
Estimated 2.1 billion as of July 2001
Organizing and categorizing document is not scalable to
the growth of internet.
Document clustering?
Is the operation of grouping similar document to
classes that can be used to obtain an analysis of the
content.
Ant clustering algorithm categorize web document to
different interest domain.
Ant Colony Models for Data Clustering
Data clustering?
is the task that seek to identify groups of similar objects
based on the value of their attributes.
 Messor sancta ants collect and pile dead corpses to form
“cemeteries” (Deneubourg et al. )

f: fraction of items in the neighborhood of the agent

k1, k2: threshold constants
 Ant Colony Models for Data Clustering

The model later extend by Lume & Faieta to include distance

function d, between data objects .
 c is a cell, N(c) is the number of adjacent cells of c, alpha is
constant
 Homogeneous Multi-agent System for
Document Clustering
 Main components: colony of agents, feature vector of web
document, 2D grid.
 Rule: agent move one step at a time to an adjacent cell. Only a
single agent and/or a single item are allowed to occupy a cell at a
time. Picking up or dropping item based on Pp & Pd
 N(c) = 8,oi is the item at cell i, g(oi) determine the similarity of oi
and other item of oj, where j E N(c)
 Density:
Homogeneous Multi-agent System for
Document Clustering
Similarity measure

r is the number of common term in doci and docj

m,n is the total number of term in doci and docj,
respectively. F is the frequency
Homogeneous Multi-agent System for
Document Clustering
Experimental Results
 Experimental data: 84 web pages from 4
different categories: Business, Computer,
Health and Science. These web page have
17,776 distinct words.
 Use 30x30 toroidal grid
 15 agents.
 tmax is 300,000. k1 and k2 in [0.01, 0.2]
increment of 0.05 for each run.
Experimental Results
 t=0
Experimental Results

 t = 50,000
Experimental Results
 t = 200,000
Experimental Results

 t = 300,000
Experimental Result

 Table
Particle Swarm Optimization

 Particle Swarm Optimization (PSO) applies to concept of social

interaction to problem solving.
 It was developed in 1995 by James Kennedy and Russ Eberhart
[Kennedy, J. and Eberhart, R. (1995). “Particle Swarm Optimization”, Proceedings of the 1995 IEEE
International Conference on Neural Networks, pp. 1942-1948, IEEE Press.] (
http://dsp.jpl.nasa.gov/members/payman/swarm/kennedy95-ijcnn.pdf )
 It has been applied successfully to a wide variety of search and
optimization problems.
 In PSO, a swarm of n individuals communicate either directly or
indirectly with one another search directions (gradients).
 PSO is a simple but powerful search technique.
Particle Swarm Optimization:
Swarm Topology

 In PSO, there have been two basic topologies used in the

literature
 Ring Topology (neighborhood of 3)

 Star Topology (global neighborhood)

I0 I0

I4 I1 I1
I4

I3 I2 I3 I2
Particle Swarm Optimization:
The Anatomy of a Particle

 A particle (individual) is composed

of: Ik
 Three vectors:
 The x-vector records the
current position (location) of the
particle in the search space, X = <xk0,xk1,…,xkn-1>
 The p-vector records the
location of the best solution P = <pk0,pk1,…,pkn-1>
found so far by the particle, and
 The v-vector contains a V = <vk0,vk1,…,vkn-1>
gradient (direction) for which
particle will travel in if x_fitness = ?
undisturbed.
 Two fitness values:
p_fitness = ?
 The x-fitness records the
fitness of the x-vector, and
 The p-fitness records the
fitness of the p-vector.
Particle Swarm Optimization:
Swarm Search

 In PSO, particles never die!

 Particles can be seen as simple agents that fly through the
search space and record (and possibly communicate) the best
solution that they have discovered.
 So the question now is, “How does a particle move from on
location in the search space to another?”
 This is done by simply adding the v-vector to the x-vector to get
another x-vector (Xi = Xi + Vi).
 Once the particle computes the new Xi it then evaluates its new
location. If x-fitness is better than p-fitness, then P i = Xi and p-
fitness = x-fitness.
Particle Swarm Optimization
25

v k  1  w  v k   20
pbest
 c1  rand ()   pbest  x  
15
 c2  rand ()   gbest  x 
v(k)
10 v(k+1)
x(k  1)  x(k )  v(k  1)  dt
5 gbest

5 10 15 20 25
Particle Swarm Optimization:
Swarm Search

 Actually, we must adjust the v-vector before adding it to the x-

vector as follows:
 vid = vid + 1*rnd()*(pid-xid)
+ 2*rnd()*(pgd-xid);
 xid = xid + vid;

 Where i is the particle,

 1,2 are learning rates governing the cognition and social
components
 Where g represents the index of the particle with the best p-
fitness, and
 Where d is the dth dimension.
Particle Swarm Optimization:
Swarm Search

 Intially the values of the velocity vectors are randomly generated

with the range [-Vmax, Vmax] where Vmax is the maximum
value that can be assigned to any vid.
Particle Swarm Optimization:
Swarm Types

 In his paper, [Kennedy, J. (1997), “The Particle Swarm: Social Adaptation of Knowledge”,
Proceedings of the 1997 International Conference on Evolutionary Computation, pp. 303-308,
IEEE Press.]
 Kennedy identifies 4 types of PSO based on 1 and 2 .
 = vid + 1*rnd()*(pid-xid)
Given: vid
+
2*rnd()*(pgd-xid);
xid = xid + vid;

 Full Model (1, 2 > 0)

 Cognition Only (1 > 0 and 2 = 0),
 Social Only (1 = 0 and 2 > 0)
 Selfless (1 = 0, 2 > 0, and g  i)
Particle Swarm Optimization:
Related Issues

 There are a number of related issues concerning PSO:

 Controlling velocities (determining the best value for Vmax),

 Swarm Size,

 Neighborhood Size,

 Updating X and Velocity Vectors,

 Robust Settings for (1 and 2),

 An Off-The-Shelf PSO

 Carlisle, A. and Dozier, G. (2001). “An Off-The-Shelf PSO”, Proceedings of the 2001 Workshop on
Particle Swarm Optimization, pp. 1-6, Indianapolis, IN. (
http://antho.huntingdon.edu/publications/Off-The-Shelf_PSO.pdf)
Particle Swarm:
Controlling Velocities

 When using PSO, it is possible for the magnitude of the

velocities to become very large.
 Performance can suffer if Vmax is inappropriately set.
 Two methods were developed for controlling the growth of
velocities:
 A dynamically adjusted inertia factor, and

 A constriction coefficient.
Particle Swarm Optimization:
The Inertia Factor

 When the inertia factor is used, the equation for updating

velocities is changed to:
 vid = *vid + 1*rnd()*(pid-xid)
+ 2*rnd()*(pgd-xid);
 Where  is initialized to 1.0 and is gradually reduced over time
(measured by cycles through the algorithm).
Particle Swarm Optimization:
Swarm and Neighborhood Size

 Concerning the swarm size for PSO, as with other ECs there is a
trade-off between solution quality and cost (in terms of function
evaluations).
 Global neighborhoods seem to be better in terms of
computational costs. The performance is similar to the ring
topology (or neighborhoods greater than 3).
 There has been little research on the effects of swarm topology
on the search behavior of PSO.
Particle Swarm Optimization:
Particle Update Methods

 There are two ways

that particles can be
I0
updated:
 Synchronously I1
I4
 Asynchronously
 Asynchronous update
allows for newly
discovered solutions to I3 I2
be used more quickly
 Te
le c
om
mu
The Future? ni c
ati
n on
s
t iPoip Me
iza e In anin g Ship di c
ur s C l e al

Se bo
a t pe
i ls

Ro
l

l f- ts
n c H u
i e tio

As
M llit an e c n En
g in

se
t e n Ma e
Sa inte

m
inte
Ma Job Schedulin

bl
nan

in
g n ce l

g
a t i o r i a
In
t a di c Da
i n a to
Mu era t E r ta m b i o n
Pe s Cl Co za t
n d ct in in
g u i m i
an g t st O pt
u er

l
e O Ch

ai
R o O so at in

M
bj psi pt u io g
Re loc
l e

s ed
ec ic im r c n
ts in
Al

em t
al e
Ve

st ibu
Sy str
Di
Dumb parts,
properly connected
into a swarm, yield
smart results.