Procesamiento Paralelo para Problemas Multiobjetivo en ...dis.um.es/~domingo/07/CD/12June/AplicacionesParalelismo/Ortega/...Reunión de Murcia, Junio de 2007 Julio Ortega Lopera Procesamiento

Reunión de Murcia, Junio de 2007

Julio Ortega Lopera

Procesamiento Paralelo para Problemas Multiobjetivo en

Entornos Dinámicos

E.T.S. Ingenierías Informática y de

Telecomunicación

Contents

Murcia, 12/06/07

1. Introduction to Dynamic Multi-objective Optimization

2. Parallel Evolutionary Multi-objective Optimization

3. PSFGA for Dynamic Optimization

4. Experimental Results

5. Conclusions and future work

Murcia, 12/06/07

• Problems which have more than one objective function(frequently in conflict) to be optimized.

• There is not one unique optimal solution, but a set ofsolutions that are not worse than any of the rest.

• This set is called the Pareto optimal set (of solutions).

Introduction: Multi-objective Problems (I)

Cost

Time 0-100Km/h

Murcia, 12/06/07

Introduction: Multi-objective Problems (II)

• Problems in which the conditions or restrictions changeover time, and so do the corresponding solutions at everyinstant.

• Most of the engineering problems are dynamic in nature.

• Examples: Scheduling, Resource allocation, Networkexploitation (telecommunications, logistics, hydraulics, gas pipelines, etc)

Introduction: Dynamic Multi-Objective Problems

Murcia, 12/06/07

The problem of finding a decision variable vector

x*(t) = [x1*(t),x2*(t), ... , xn*(t)] ∈Rn

satisfying a given restriction set

g(x,t) ≤ 0, h(x,t) = 0

and optimizing the function vector:

f(x,t) = {f1(x,t), f2(x,t),…, fm(x,t)}

Murcia, 12/06/07

Introduction: Dynamic Multi-Objective Problems

Multi-objective Evolutionary Algorithms– Average funtions,– VEGA (1984), Schaffer– MOGA (1993), Fonseca– NSGA-II (2000), Kalyanmoy Deb– SPEA2 (2001), Eckart Zitzler

EvoDOP: http://www.aifb.uni-karlsruhe.de/mailman/listinfo/evodop

Parallel Processing- MOGADES (2002), Kamiura- Parallel NSGA-II (2004), Sergio Nesmachnow- PSFGA (2004):

• F. Toro, J. Ortega, E. Ros, S. Mota, B. Paechter, J. M. Martín, PSFGA: Parallel Processing and Evolutionary Computation For Multiobjective Optimisation, Parallel Computing 30 (5–6) (2004) 721–739.

• F. Toro, E. Ros, S. Mota, J. Ortega, Evolutionary Algorithms for Multiobjective and Multimodal Optimization of Diagnostic Schemes, IEEE Trans. on Biom Eng.53 (2006)

Introduction: previous works

Dynamic Multi-Objective Problems: Example (I)

Murcia, 12/06/07

Poblem of Control of Water Supply Network: distributing the flow that the water treatment station (ETAP), such that demand is satisfied, the tanks neither overflow nor fall below an established minimum water volume, and minimize valve movements.

Dynamic Multi-Objective Problems: Example (II)

Murcia, 12/06/07

Prediction OptimizationProcedure NDPA

ModelInverse Valves

Valves

Neural Aprox.

Tanks

W’(t+1)

X(t)

Parameters

Y’(t+1)

P(t+1)Y(t+1)

W(t+1)X(t+1)

Using ANNs(RBF)

X(t) Levels of the tanks.W(t) Predicted demand volumes.P(t) Valve position consigns.Y(t) Tank inputs volumes.

Dynamic Multi-Objective Problems: Example (III)

Murcia, 12/06/07

• With predicted input vectors module simulation can estimate sets of feasible control values for the following stages, and generate sample trajectories in order to approximate the cost-to-go function at stage i+1:

i i+1 i+2 i+3 N. . .

Y(i+1)1

Y(i+1)2

Y(i+1)M i+1

j1 j2 j3 jN-i

End of the horizon

Future evolution of the system

Contents

1. Introduction to Dynamic Multi-objective Optimization

2. Parallel Evolutionary Multi-objective Optimization

3. PSFGA for Dynamic Optimization

4. Experimental Results

5. Conclusions and future work

Murcia, 12/06/07

Parallel Processing: Expected Benefits

It is expected to get better performance

Regarding to the multi-objective problem:

1. Solutions lying nearer to the actual Pareto front, and2. More diversified solutions.

Regarding to the dynamic problem:

1. To get solutions suitable for the new problem conditions, before they will be needed, and

2. To be able to meet more restrictive time limits, by meansof increasing the necessary resources for the solutionsprocessing.

3. Diversity control (after changes and during execution)

Murcia, 12/06/07

Murcia, 12/06/07

W(t) o X(t)W(t) o X(t)

• Improve the performance of the control procedure as more processors:

0

500

1000

1500

2000

2500

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Y(t)

0

500

1000

1500

2000

2500

1 3 5 7 9 11 13 15 17 19 21 23

Y(t)

0

3000

6000

9000

12000

15000

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

X(t)

0

3000

6000

9000

12000

15000

1 3 5 7 9 11 13 15 17 19 21 23

X(t)

0

500

1000

1500

2000

2500

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Y(t)

0

500

1000

1500

2000

2500

1 3 5 7 9 11 13 15 17 19 21 23

Y(t)

0

3000

6000

9000

12000

15000

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

X(t)

0

3000

6000

9000

12000

15000

1 3 5 7 9 11 13 15 17 19 21 23

X(t)

Results using only one of the processors Results using eight processors

Parallel Processing: Example (I)

Murcia, 12/06/07

Parallel Processing: Example (II)

• Computing times and the speedups obtained by the parallel implementation of the procedure, using the PVM message-passing library in a cluster of 9 PCs (Pentium II, 333 MHz) with FastEthernet.

Speedup obtained for 9 trajectories

0123456789

1 2 3 5 9

Spee

dups

20,1 min.11,1 min.7,8 min.8 Proc.

48,5 min.25,6 min.13,9 min.4 Proc.

145 min.74,7 min.42,0 min.1 Proc.

32 Traj.16 Traj.8 Traj.Proc./ Traj.

20,1 min.11,1 min.7,8 min.8 Proc.

48,5 min.25,6 min.13,9 min.4 Proc.

145 min.74,7 min.42,0 min.1 Proc.

32 Traj.16 Traj.8 Traj.Proc./ Traj.

Proc.

Efficiency obtained: 0.8 – 0.9

Computing times for different number of trajectories and processors

Contents

1. Introduction to Dynamic Multi-objective Optimization

2. Parallel Evolutionary Multi-objective Optimization

3. PSFGA for Dynamic Optimization

4. Experimental Results

5. Conclusions and future work

Murcia, 12/06/07

Description of the PSFGA Algorithm (I)

• In PSFGA, a master process generates subpopulationswhich are distributed among the worker processes,

• The workers will run the EA on the subpopulationsduring genpar iterations,

• After genpar iterations, the master receives thesubpopulations, runs the algorithm over the wholepopulation (genser iterations) and selects the newpopulation to be distributed among the workers or to giveas results to the problem,

• Selection is carried out from the non-dominatedsolutions which after crossover and mutation willgenerate the offspring.

Murcia, 12/06/07

Description of the PSFGA Algorithm (II)

Murcia, 12/06/07

….

Worker 1 (N/P indiv.)

Worker 2 (N/P indiv.)

Worker P-1 (N/P indiv.)

Worker P (N/P indiv.)

Master (N/P indiv.)

genseriterations genpar

iterations

f2

f1

P1P2 P3

P4

P5

Description of the PSFGA Algorithm (III)

Murcia, 12/06/07

Selection

Transf.

+

Non-dominatedindividuals

Population

Crossover (1P)Gaussianuniformmutation

Parentspopulation

Offspringpopulation

Crowding AlgorithmThreshold

Radius

F2

F1Frente de Pareto

Not eliminated solutions

Eliminated solutions

Contents

1. Introduction to Dynamic Multi-objective Optimization

2. Parallel Evolutionary Multi-objective Optimization

3. PSFGA for Dynamic Optimization

4. Experimental Results

5. Conclusions and future work

Murcia, 12/06/07

Evaluation of the Algorithm

• Accuracy

• Stability

• Reaction

) (t)V (t),V ,)t(V of terms (in solutionthe ofquality :acc(t) minmax

acc(t)})acc(t,{stb(t) −−= 10max

t}-{Mgen )-(1

acc(t))acc(t'Mgent't

such that min{t'-t )reac(t, ∪⎪⎩

⎪⎨

⎧

⎪⎭

⎪⎬⎫

≥

≤<=

εε

Murcia, 12/06/07

Experimental Results: Test Problems

FDA1 FDA2

12 1 ff −=

Murcia, 12/06/07

τ intervalo entre cambios

Experimental Results: Speedup (I)

τi=15

τi=5

τi=25

Murcia, 12/06/07

FDA2

Experimental Results: Speedup (II)

[ ][ ] )P,M(O)t

PMB()t

PMA(genpar)tMB()tMA(genser

)tMB()tMA(gen)P(S

1

r

01r

0

1r

0

+⎥⎥⎦

⎤

⎢⎢⎣

⎡×⎟

⎠⎞

⎜⎝⎛×+×⎟

⎠⎞

⎜⎝⎛××+××+×××

××+×××=

Murcia, 12/06/07

Ganancias Superlineales:

genser+genpar<gen

Mejores condiciones de diversidad y necesita menos iteraciones para alcanzar soluciones similares

Experimental Results: Speedup (III)

Murcia, 12/06/07

Experimental Results: Solution Quality

τi Hiperv.

(± 0.005) acc

(± 0.01) stb (± 0.001)

reac (ε = 0.1)

5 0.345 0.97 0.0 5 10 0.350 0.99 0.0 5 15 0.355 1.00 0.0 5 20 0.347 0.98 0.020 5 25 0.352 0.99 0.0 5 30 0.347 0.98 0.014 5 35 0.347 0.98 0.0 5 40 0.344 0.97 0.010 5 45 0.349 0.98 0.00 5 50 0.345 0.97 0.009 5

Murcia, 12/06/07

FDA2

FDA1

Contents

1. Introduction to Dynamic Multi-objective Optimization

2. Parallel Evolutionary Multi-objective Optimization

3. PSFGA for Dynamic Optimization

4. Experimental Results

5. Conclusions and future work

Murcia, 12/06/07

Conclusions and Future Work

• Parallel processing:

– New schemes to distribute all or part of the selectionprocess carried out by the master among the workers.

– To study the way parallelism can improve diversity andconvergence (emergent properties)

• Algorithm:

– To add more test problems and real applications.

– To study new adaptation and prediction strategies.

Murcia, 12/06/07

References• Damas, M.; Salmerón, M.; Ortega, J.; Olivares, G., Pomares, H.:”Parallel

Dynamic Water Supply Scheduling in a Cluster of Computers”. Concurrency: Practice and Experience, Vol.13, pp.1281-1302, 2001

• F. Toro, J. Ortega, E. Ros, S. Mota, B. Paechter, J. M. Martín, “PSFGA: Parallel Processing and Evolutionary Computation For MultiobjectiveOptimisation”, Parallel Computing 30 (5–6) (2004) 721–739.

• F. Toro, E. Ros, S. Mota, J. Ortega, “Evolutionary Algorithms for Multiobjectiveand Multimodal Optimization of Diagnostic Schemes”, IEEE Trans. on BiomEng.53 (2006)

• Cámara, M.; Ortega, J.; Toro, F.J.:”Parallel Processing for Multi-Objective Optimization in Dynamic Environments”. Workshop on NIDISC, IPDPS’07, 2007.

• Cámara, M.; Ortega. J.; Toro, F.J.:”The Parallel Single Front Genetic Algorithm (PSFGA) in Dynamic Multi-Objective Optimization”. International Workshop in Artificial Neural Networks (IWANN’07), 2007.

Murcia, 12/06/07