Artificial Systems 2

8/2/2019 Artificial Systems 2

1/47

Adaptive Systems

Lecture 9: Artificial Adaptive Systems (2)

Dr Giovanna Di Marzo Serugendo

Department of Computer Science

and Information Systems

Birkbeck College, University of London

Email: [email protected]

Web Page: http://www.dcs.bbk.ac.uk/~dimarzo
http://www.dcs.bbk.ac.uk/~dimarzohttp://www.dcs.bbk.ac.uk/~dimarzohttp://www.dcs.bbk.ac.uk/~dimarzohttp://www.dcs.bbk.ac.uk/~dimarzo


2/47

Adaptive Systems - Spring 2008 - Giovanna Di Marzo Serugendo 2

Lecture 8: Review

Taxonomy / Classification

Static Optimisation Problems

Ant-Colony Optimisation Particle Swarm Optimisation

Dynamic Optimisation Problems

Trust-based access control


3/47

Adaptive Systems - Giovanna Di Marzo Serugendo 3

Lecture 9: Overview

Swarms

Robots

Spiders-based systems

Manufacturing Control

Immune Computer

P2P Systems Autonomic Computing


4/47


Swarms of Robots

Cooperative prey transport by social insects Ants recruit other ants for collaborative transport of

preys too heavy to be carried by a single ant E.g. 100 ants transporting a worm (5000 times bigger than

each single ant)

Resistance to traction decides ants to recruitnestmates

Size of group is adapted to size of prey Pheromone used to recruit nestmates

Coordination for transporting prey occurs throughindirect communication (stigmergy) Actual transport involves:

re-alignment and re-positioning


5/47


Swarms of Robots

Application Swarms of robots [Bonabeau 99]

Collaborative box-pushing Indirect communication Decentralised control

Goal: Localise a box in a given space and push it towards an edge

Subsumption architecture: Every behaviour is subdivided into atomic sub-behaviours activated

when necessary (reactive approach) Each sub-behaviour has its own sensors inputs and actuators

outputs Hierarchy of behaviours with priority Arbitration module controls actual activation of sub-behaviour


6/47


Swarms of Robots

Sensors Left/Right infrared (obstacles) and photocells sensors (box)

Steering actuator

Left/Right wheel motors

Behaviours definition Find (box) lowest priority

Follow (other robot)

Slow (neighbour collision)

Goal (move towards box)

Avoid (obstacle collision change direction) highest priority


7/47


Swarms of Robots

Scenario

Goal activated

Follow and Goal: set motion

Avoid: stop current process (Goal deactivated) ->re-alignment and re-positioning

Goal of one robot is re-activated

No direct communication (stigmergy)

Robots implementation

Model and simulation of ants prey transports


8/47


Swarms of Robots

Adaptation

Different configurations

Box positioning, robots placements


9/47


Region Detection

Metaphor: Social Spiders

Few species of spiders are social

Sharing of web

Collaboration (preys, web weaving)

Stigmergy based on silk

Spiders follow silk or move to points where silk is fixed


10/47


Region Detection

http://media.star-telegram.com/Multimedia/News/Photos/Bigweb.jpg


11/47


Region detection

Region detection (grey levels) [Bourjot 03]

Partition of image into subsets of separate objects

Determination of sets of connected pixels (regions)

Idea:

Webs weaving determines the region

Algorithm

Spider has to detect a given region (grey level)

Several spiders explore image and fix silk on relevant pixels

Silk attraction

Resulting web is fixed on interesting pixels


12/47


Region Detection


13/47


Region Detection


14/47



Manufacturing control Management of internal logistic and of production system

Routing of product instances

Assignment of workers

Assignment of raw material Operations begin and end

Dynamic environment Failures, new products, equipment upgrades, etc.

Idea: Decentralised control with self-organising behaviour


15/47



Metaphor: Ant foraging

PROSA Architecture [Hadeli 03]

Agents:

Orders agents (logistics for managing products), products agents(processes tasks), resources agents (raw material, machines, etc)

Mapping of control and production system into agents Actual production system is reflected into an agents structure Each resource/product/order has a corresponding

resource/product/order agent (local information only) Links among agents (e.g. order agents know about location of

resources agents to products agents necessary to complete order) Agents creates ant-agents (mobile agents) that explore the cyber

production system and deposit/sense pheromone


16/47



Ant-agents behaviour Feasibility information ants

Information related to the resource locations (availability, speed,etc.)

Exploring ants

Order agents create several ants each exploring a way of realisingthe order (gives back a report with followed route)

Intention propagation ants Order agents create ants that propagate information about the

orders intentions (chosen best route). Ant has a fixed route, andmakes bookings.

Manufacturing control Obtained from the choices made by order agents

On the basis of the above information

Actually executed by resources agents


17/47



Exploring Ants (EA)

Tries to find solutions

Searching for solutions is guided by local pheromones

Reports result of solution to the corresponding Order Agent

time

Orders - Agent

Ra

Rb

Rc

Rd

t1

EA1

Ra

Rb

Rc

Rd

t2

EA2

Ra

Rb

Rc

Rd

t3

EA3


18/47



Adaptation

Actual factory status

Current orders


19/47

Artificial Immune System

Design of an artificial immune system

Representation for the components of the system

Mechanisms to evaluate the interaction of

individuals with environment and with each other Procedures of adaptation dynamics of system

Used for:

Modelling immune systems

Solving problems using artificial immune systems


20/47


Representation Abstract model of immune cells and molecules

Recognition of antigen by cell (antibody) receptor

Occurs through shape complementarity or shape similarity

Model of shape recognition

Data structure: attribute string

Real-valued vector / Integer vector / ...

Shape recognition is based on:

Similarity/affinity measure between attribute strings ofantigen and antibody

Ab = (Ab1, ..., Ab

L) Ag = (Ag

1, ..., Ag

L)

Affinity D: SL x SL RL (degree of matching)


21/47


Evaluating interactions

Affinity measure Affinity D: SL x SL RL

Complementary / Similarity A distance: (Euclidian, Manhattan, Hamming)

http://en.wikipedia.org/wiki/Immune_system

Ab = [1 0 0 0 1 1 0 0 1]

Ag = [1 1 0 0 0 1 0 1 0]

Match(Ab, Ag) = 0 1 0 0 1 0 0 1 1

Complementarity: affinity = 4 (how different)Similarity: affinity = 5 (how similar)

L

i iiAgAbD

1

2)(

L

i iiAgAbD

1)(

http://en.wikipedia.org/wiki/Immune_system


22/47

Artificial Immune Systems

Immune Algorithms

Bone marrow

Generate populations of immune cells and molecules (tobe used in artificial immune system)

Negative selection

Learning phase (avoid matching self)

Define set of detectors (for anomaly detection)

Clonal selection

Generate additional immune cells driven by detectedantigens

Immune networks

Simulate immune networks


23/47


Bone Marrow

Generation of antibodies

Model 1: Generation of attribute string of length L with random values

Model 2 Generation of antibodies from gene library

Concatenation of genes from gene library


24/47


Negative Selection Algorithms T cells mature in Thymus

Learn to distinguish self from non-self

T cells that cannot distinguish self properly must be destroyed

Model Create T cells bit strings of length L

Test T cells against known set of self-patterns (S)

Discard the ones that match the self-patterns (affinity measure)

Otherwise allow T cell to enter set of detectors

Monitoring/Protection:

test detectors against set of strings to protect If matching then an anomaly (non-self) has been detected

Applications

Computational security


25/47

Artificial Immune Systems

Clonal selection

Proliferation of cells that recognise specificantigen

Proportional to degree of affinity (higheraffinity, higher proliferation)


26/47


Intrusion Detection

Metaphor: Mammalian Immune System (Lecture 2)

Self/Non-Self recognition Each cell has a marker (self) Cells without marker (non-self) are considered antigen

Immune system attacks antigen

Agents of Immune System B cells - Detection

Wait for antigen, replicate and release antibodies Antibodies mark antigen (intruders)

T cells - Response Destroy marked cells B and T cells

Transported by blood and lymphatic vessels across the whole body


27/47


Intrusion Detection

Characteristics of Immune System

Robust

Decentralised and distributed (no central control)

Dynamic (new components created, destroyed, circulated)

Tolerant to errors (failure of single components has a minimalimpact)

Adaptable

Learn to recognise new infections

Memory of past infections

Autonomous

No outside control


28/47


Intrusion Detection

Intrusion Detection - ARTIS [Forrest 99] ARTificial Immune System

(Mobile) detectors circulating in the system Stand for the T, B cells and antibodies

Detection Bit strings stand for proteins to detect Random generation of detectors (random string)

Look for matching portions of strings (anomaly)

Training:

If detector matches a self string then detector is destroyed and new oneregenerated

(Associative) Memory Mapping of identified non-self strings to responses


29/47


Intrusion Detection

Applied to: Computer virus detection Host-based intrusion detection Network intrusion detection

Proteins = network traffic Strings =

(Source IP address, Destination IP address, TCP Port Service) Anomaly detection = high frequency of connections

Environment Network of computers

Each computer runs a detector node Experiments

Off-line with actual data, no mobile detectors Detection of attacks


30/47


Intrusion Detection

Adaptation

Learning

Memory:

Quick reaction for further identical intrusion

Adaptation to changes in normal behaviour


31/47


Network Intrusion Detection and Response

Combination of Metaphors [Foukia 05]: Intrusion Detection

Metaphor: Immune System Implementation: mobile agents

Anomaly: bad sequence of events

Alert: triggers the diffusion of pheromone

Intrusion Response

Metaphor: Ants foraging Implementation: mobile agents

Mobile agents trace the source of the attack (machine) as ants follow atrail for food

Mobile Agents Software able to change its location (keeping its execution state)


32/47

P2P / Networks

T-Man Algorithm [Jelasity 05]. Generic protocol based on gossip communication model Goal: network topology management problem

Nodes randomly connected

Re-organisation of connections to produce desirable topology Nodes become neighbours based on information such as:

geographic position, content, storage capacity

Metaphor: Gossip Periodic exchange and update of information among

members of a group Allows: aggregation of global information inside a population,

social learning Parameters: neighbourhood, level of precision of information


33/47

P2P / Networks

Principle Nodes maintain local list of of (logical) neighbours a fixed

number of neighbours, say c For each neighbour a profile is stored

Profile is relevant for the topology to achieve (type of data stored, ID,location, etc)

Ranking function defines the target topology (e.g. distance) Serves for reorganising the set of neighbours Based on profile of the nodes (distance between profiles)

Gossip message exchange Choice of closest neighbour based on ranking function Local exchange / combination of neighbours profile

Merging of neighbours profile Keep c closest neighbours Drop the rest

Nodes become closer and closer

Allows adaptation of neighbours list Re-organisation of the network topology


34/47

P2P / Networks

Applications

Overlay networks supporting P2P systems

Maintenance or establishment of P2P topology

Sorting, Clustering, Distributed Hash table


35/47

Autonomic Computing

Autonomic computing

computing systems that can

manage themselvesgivenhigh-level objectivesfrom

administrators [Kephart03]

35


36/47


Autonomic Computing

Metaphor: human nervous system Regulation of vital functions:

Breath, blood pressure, heart beating, Seamlessly for human being

Autonomic Computing Goal:

Machines that manages themselves (self-management) With highest performances 24/7

Human Nervous System metaphor Not used for implementation!

Artificial mechanisms employed


37/47


Autonomic Computing

Self-Configuration (installation, configuration, integration)

Automated configuration of components and systems followhigh-level policies. Rest of System adjusts automatically and

seamlessly [Kephart03]

Self-Optimisation (parameters)

Components and systems continually seek opportunities toimprove their own performance and efficiency [Kephart03]


38/47


Autonomic Computing

Self-Healing (error detection, diagnostic, repair)

System automatically detects, diagnoses, and repairs localizedsoftware and hardware problems [Kephart 03]

Self-protection (detection and response to attacks)

System automatically defends against malicious attacks or

cascading failures. It uses early warning to anticipate andprevent system wide failures [Kephart 03]


39/47


Autonomic Element

Autonomic Manager

Managed Resource


40/47


Autonomic Elements


41/47


Example

Two Applications Managers (AM1, AM2) handlingresources (servers S1, S2)

Resources are dynamically allocated on the basis ofpolicies

If application manager cannot apply its policy, it asks aResource Arbiter (RA) for additional resources

RA

AM1 AM2

S1 S2


42/47


Example

Components

Two application managers (AM1, AM2)

Resource Arbiter (RA)

Two Servers (S1, S2)

Meta-data

Servers Transaction Time

Servers CPU availability


43/47


Example

Policies

AM Policy1 Increase CPU by 5% if response time is above 100ms

AM Policy 2 If transaction time > 100 ms and CPU availability > 98%, ask

RA for more CPU

RA Policy

If request for CPU, grant and give priority to AM1


44/47


Unity

Autonomic Elements

Application Environment Manager (AM)

Management of environment resources and Communications

Prediction about impact in increasing/decreasing resources

Utility function

Resource Arbiter (RA)

Allocation of resources

Computation of Optimum

Resources (servers)

Registry: Location of autonomic elements Registry policy: High-level policies (utility function)

Sentinel: Monitors elements for another element

[Chess 04]

),(iii DSU

i iii DSU ),(


45/47


Unity


46/47


Readings

[Bonabeau 99] E. Bonabeau, M. Dorigo, and G. Thraulaz. SwarmIntelligence: From Natural to Artificial Systems Santa Fe Institute Studies onthe Sciences of Complexity. Oxford University Press, UK, 1999.

[Hadeli 03] Hadeli et al. Self-organising in Multi-Agent Coordination and

Control using Stigmergy. LNAI 2977, Springer, pp. 105-123, 2004.

[Foukia 05] N. Foukia: IDReAM: Intrusion Detection and Response

executed with Agent MobilityArchitecture and Implementation.AAMAS05, 2005.

[Hofmeyr 99] S. A. Hofmeyr, S. Forrest: Architecture for an Artificial

Immune System. Evolutionary Computation 7(1):45-68, 1999.


47/47


Readings

[Bourjot 03] Bourjot, C. Chevrier, V. and Thomas, V.:A new swarmmechanism based on social spiders colonies: from web weaving toregion detection. In Web Intelligence and Agent Systems: AnInternational Journal - Vol 1, N.1, pp 47-64 WIAS. 2003.

[Chess 04] Chess et al. Unity: Experiences with a prototype

Autonomic Computing System. ICAC'04. 2004.

[Jelasity 05] Mrk Jelasity and Ozalp Babaoglu: T-Man: Gossip-based overlay topology management. In Proceedings ofEngineering Self-Organising Applications (ESOA'05), July 2005.

[Kephart 03] J. Kephart, D. Chess: The Vision of AutonomicComputing, IEEE Computer, January 2003, 36(1):41-50, 2003.

Documents

Artificial Systems 2