02 lgambardella ppt
TRANSCRIPT
-
8/3/2019 02 LGambardella PPT
1/30
Ant Colony Optimization for ad-hoc networks
The First MICS Workshop on Routing for Mobile Ad-Hoc Networks
Zurich, February 13, 2003, 14:00-17:00, ETH-Room ETZ E8
Luca Maria Gambardella
IDSIA, Lugano
Universit della Svizzera italiana
Scuola universitaria professionale
della Svizzera italiana
IDSIA
Istituto Dalle Molle di studi
sullintelligenza artificiale
-
8/3/2019 02 LGambardella PPT
2/30
Ant Colony Optimization Algorithms
Ant Colony Optimization Algorithms
ACO algorithms are inspired by the observation of realants (Dorigo, Colorni, Maniezzo, 1991)
Real Ants are social insects organized in colonies.
Ant colonies show a high structural level compared to thesimplicity of the single individual
Ants coordinate their activities by an indirect form ofcommunication (stigmergy) based on pheromone laying
Foraging behavior: searching for food by parallelexploration of the environment
-
8/3/2019 02 LGambardella PPT
3/30
First they explore
Individual ants mark their path by emitting achemical substance - a pheromone - as they forage for food
Ants smell pheromone and they tend to choose path with strong pheromoneconcentration
Other ants use the pheromone to find the food source
When the system is interrupted, the ants are ableto adapt by rapidly adopting second best solutions
Social insects, following simple, individual rules, accomplish complex colonyactivities through: flexibility, robustness and self-organization
How Ants Find Food
How Ants Find Food
-
8/3/2019 02 LGambardella PPT
4/30
Pheromone Trail Following
Pheromone Trail Following
Ants and termites follow pheromone trails
-
8/3/2019 02 LGambardella PPT
5/30
Asymmetric Bridge ExperimentAsymmetric Bridge Experiment
Goss et al., 1989
-
8/3/2019 02 LGambardella PPT
6/30
ClientsRequests
Time Windows
Pick-up and deliveryAccess Limitation
FleetNon-homogeneous vehicles
Costs (trucks own/external)DriversTime limitation
InformationDriving time
Limitation on max km.Depots, number, location
ACO for routing
in Transport
Problems
-
8/3/2019 02 LGambardella PPT
7/30
colony
food
nest
pheromone
fleet
goods
depot
OptimizedSolution
-
8/3/2019 02 LGambardella PPT
8/30
Given a transport
optimization problem
ACO for transport problems3FNSs 1994-2003, CTI 2000/2, EU5 Metanet 2001/4, EU5 Mosca 2002/3 )
Lugano
Mendrisio
Cost km
ArtificialPheromone
Ants probabilisticallybuild solutions using costand pheromone
Best solutions arerewarded with newpheromone
The Algorithm learns from experience tocompute always better solutions
-
8/3/2019 02 LGambardella PPT
9/30
IDSIA Algorithms best-known in literature in their domain (2002)
MACS-VRPTW: Vehicle Routing Problem with Time Windows
Gambardella L.M, Taillard E., Agazzi G., MACS-VRPTW: A Multiple Ant Colony System for Vehicle
Routing Problems with Time Windows , In D. Corne, M. Dorigo and F. Glover, editors , New Ideas inOptimization. McGraw-Hill, UK, pp. 63-76, 1999
(best of 8 algorithms over 100 test problems,16 new best known results)
HAS-QAP: Quadratic Assignment ProblemGambardella L.M, Taillard E., Dorigo M., Ant colonies for the Quadratic Assignment Problem ,
Journal of the Operational Research, Society, 50, pp.167-176, 1999
(best known for structured problems)
HAS-SOP: Sequential Ordering ProblemGambardella L.M, Dorigo M., An Ant Colony System Hybridized with a New Local Search for theSequential Ordering Problem, INFORMS Journal on Computing, vol.12(3), pp. 237-255, 2000
(22 new best known results on 33 test problems)
Flexible Job Shop ProblemMastrolilli, M., Gambardella, L.M., Effective Neighborhood Functions for the Flexible Job ShopProblem, Journal of Scheduling, 1, 2000, vol. 3(1), pp. 3-20, 2000.
(116 results improved on 221 problems, 10 time faster than the previous best)
-
8/3/2019 02 LGambardella PPT
10/30
CADIS-OPT-FLEET (2001-2002)
AntOptima IDSIA MIGROSTours optimization in all Switzerland for non-food goods
from the new hub center in Suhr (150-200 vehicles)
Non-homogeneous trucksTrailers
Soft TWHub use restrictions
Tours MinimizationDistance Minimization
Integration with CADIS
-
8/3/2019 02 LGambardella PPT
11/30
AntNet Applied toRouting in Internet-like Networks
AntNet Applied toRouting in Internet-like Networks
?
Probabilistic rule tochoose the path
Pheromone traildepositing
Di Caro and Dorigo, 1997
Source
Destination
Memory
-
8/3/2019 02 LGambardella PPT
12/30
AntNet is an instance of an ACO algorithm for:distributed andadaptive routing in communications networks
In distributed adaptive routing at each network node theroutingpolicyis continually adapted to the variations in the input traffic
patterns
A routing policy is a local mapping parametrized by a data
structure calledrouting table
It is assumed that a robust model of the input traffic is notavailable: the policy should be learned
AntNetAntNet
-
8/3/2019 02 LGambardella PPT
13/30
The Routing ProblemThe Routing Problem
The practical goal of routing algorithms is to buildrouting tables
Routing is difficult because costs are dynamic Adaptive routing is difficult because changes in the
control policy determine changes in the costs andvice versa
Destination
node j
Routing table of node k
Next node ij
...
...
...
...
1
i1
N
iN
k-1
ik-1
...
...
k+1
ik+1
-
8/3/2019 02 LGambardella PPT
14/30
AntNet: The AlgorithmAntNet: The Algorithm
Ants are launched at regular instants from each node torandomly chosen destinations
Ants are routed probabilistically with a probability function of:(i) someartificial pheromone values, and(ii) someheuristic values, maintained on the nodes
Antsmemorize visited nodes and elapsed times Once reached their destination nodes, ants retrace their
paths backwards, and update the routing tables
AntNetis distributed and not synchronized
-
8/3/2019 02 LGambardella PPT
15/30
The AntNet data structureThe AntNet data structure
-
8/3/2019 02 LGambardella PPT
16/30
Forward AntForward Ant
-
8/3/2019 02 LGambardella PPT
17/30
Backward AntBackward Ant
-
8/3/2019 02 LGambardella PPT
18/30
Ants Pheromone Trail DepositingAnts Pheromone Trail Depositing
ijd
kt + 1( ) 1 ( ) ijd
kt( ) + ijd
kt( )
where the (i,j)s are the links visited by ant k,
and
ijd
kt( ) = qualityk
where qualityk is set proportionalto
the inverse of the time it tooks ant kto build the path from ito dviaj
i
j
tijd
Source
Destination
d
-
8/3/2019 02 LGambardella PPT
19/30
AntNet: Experimental setupAntNet: Experimental setup
Realistic simulator (though not industrial) Many topologies Many traffic patterns Comparison with many state-of-the-art algorithms
(Open Shortest Path First, SPF Straight Packet Forwarding, AdaptiveBellman-Ford, Q-routing, Predictive Q-routing, Ideal Deamon)
Performance measures:throughput (bit/sec) measures the quantity of service, and averagepacket delay (sec) measures the quality of service
Japanese NTT net
American NSF net
-
8/3/2019 02 LGambardella PPT
20/30
AntNet: Some Results (1)AntNet: Some Results (1)
NSF net NTT net
Increasing traffictraffic increased by reducing the mean session inter arrival time
Throughput(b/s)
Avg
packet
delay(s)
From Di Caro and Dorigo, 1998,Journal of Artificial Intelligence
Research
-
8/3/2019 02 LGambardella PPT
21/30
AntNet: Some Results (2)AntNet: Some Results (2)
Data averaged over a 5 seconds sliding window
NSF net NTT net
From Di Caro and Dorigo, 1998,Journal of Artificial Intelligence
Research
-
8/3/2019 02 LGambardella PPT
22/30
Research Lines (1)Research Lines (1)
AntNetMobile: a pheromone based routing algorithm basedon AntNet for mobile ad hoc networks.
Reactive vs Proactive routing
We are investigating a mixed approach where updatedinformation are maintained for few nodes only. New pathsare discovered on demand by using the pheromoneinformation
We expect to be able to maintain multiple disjoint routes.One of the goal is the identification and the management ofcritical links
-
8/3/2019 02 LGambardella PPT
23/30
We are evaluating the overhead of sending aroundartificial ants.
Two possible ways to reduce overhead are piggyback
information on standard data package or move somefunctionality to the MAC level
AntNet uses time as measure for path performance. Thisrequires clock synchronization between nodes.
We are investigating the use of the number of hops orpower consumption.
Research Lines (2)Research Lines (2)
-
8/3/2019 02 LGambardella PPT
24/30
QoS negotiation
Combining routing with other functions like searching
or discovery
Completely distributed or hierarchical organization?.
Bio-inspired pheromone based cluster of nodes
Research Lines (3)Research Lines (3)
-
8/3/2019 02 LGambardella PPT
25/30
Bison: Biology-Inspired Techniques forSelf Organization in Dynamic Networks
Bison: Biology-Inspired Techniques forSelf Organization in Dynamic Networks
Bison is a three years European FET project startingJanuary 2003.
Partners:
University of Bologna, Italy (Coordinator)
Telenor Communication, Norway
Technische Universitaet Dresden, Germany
IDSIA, Switzerland
Santa Fe Institute, USA
-
8/3/2019 02 LGambardella PPT
26/30
Bison: GoalsBison: Goals
BISON explores ideas derived from complex adaptive systems (e.g.Ant Colony Optimization) in highly dynamic network environments.
Systems: Ad Hoc networks, P2P and Grid computing
Themes: Flexible, adaptive, and decentralizednetwork functions
Routing, searching, discovery, document sharing, load-balancing
New simulation environment
-
8/3/2019 02 LGambardella PPT
27/30
Failure location, restoration andresilience in optical and ad-hoc networks
Failure location, restoration andresilience in optical and ad-hoc networks
Is a three years Swiss Hasler Foundation project started inNovember 2002.
Partners:
EPFL I&C-IIF, DIE Lugano
Critical links
Repairing and reroute in case of failure
Dynamic computation of alternative paths
Dynamic clustering
MOBILE MAN
FET EU Projectof 3 years
Partners
CNR Pisa
DIE/DLS Lugano
Uni Cambridge
Eurocom France
Uni Helsinki
-
8/3/2019 02 LGambardella PPT
28/30
SWARM-BOTS EU5 (2001-2004)IRIDIA, EPFL, CNR-IP, IDSIA
Collective robotics inspired by colony behaviour
Guy Theraulaz
-
8/3/2019 02 LGambardella PPT
29/30
Ant Colony Optimization Major PublicationsAnt Colony Optimization Major Publications
-
8/3/2019 02 LGambardella PPT
30/30
Senior researchers PhD. Students
Gianni di Caro Leonora BianchiAlberto Donati Frederick DucatelleIvo Kwee
Monaldo Mastrolilli Sofware engineersRoberto Montemanni Norman CasagrandeGiovanni Pettinaro David Huber
Andrea Rizzoli Fabrizio Oliverio
IDSIA people involved in ACO ProjectsIDSIA people involved in ACO Projects