Dynamic Optimization using Self-Adaptive Differential Evolution

Transcription

1 Dynamic Optimization using Self-Adaptive Differential Evolution IEEE Congress on Evolutionary Computation (IEEE CEC 2009), Trondheim, Norway, May 18-21, 2009 J. Brest, A. Zamuda, B. Bošković, M. S. Maučec, V. Žumer Faculty of Electrical Engineering and Computer Science University of Maribor May 19, 2009 Dynamic Optimization using Self-Adaptive Differential Evolution 1 / 29

2 1 Introduction 2 Background The Differential Evolution Algorithm The Self-adaptive DE Algorithm 3 Algorithm for DOPs 4 Experimental Results 5 Conclusions Dynamic Optimization using Self-Adaptive Differential Evolution 2 / 29

3 Introduction Differential Evolution (DE) is simple yet powerful EA algorithm for optimizing continuous functions, e.g. static optimization environment. CEC 2009 special session on evolutionary computation in dynamic and uncertain environments. Main goal: self-adaptive DE algorithm + multi-populations Dynamic Optimization using Self-Adaptive Differential Evolution 3 / 29

4 Introduction In this presentation: hybridization of our self-adaptive differential evolution algorithm jde with multi-populations, aging, overlapping search performance comparison on the set of benchmark problems Dynamic Optimization using Self-Adaptive Differential Evolution 4 / 29

5 The Differential Evolution Algorithm 1 Introduction 2 Background The Differential Evolution Algorithm The Self-adaptive DE Algorithm 3 Algorithm for DOPs 4 Experimental Results 5 Conclusions Dynamic Optimization using Self-Adaptive Differential Evolution 5 / 29

6 The Differential Evolution Algorithm The Differential Evolution Algorithm Price & Storn, 1995 (JGO:1997) NP.. population size (D-dimensional vectors) F.. mutation scale factor CR.. crossover parameter Dynamic Optimization using Self-Adaptive Differential Evolution 6 / 29

7 The Differential Evolution Algorithm The Differential Evolution Algorithm Price & Storn, 1995 (JGO:1997) NP.. population size (D-dimensional vectors) F.. mutation scale factor CR.. crossover parameter rand/1 strategy: v i (G) = x (G) r 1 + F ( x r2 (G) x r3 (G) ), r 1 r 2 r 3 i Dynamic Optimization using Self-Adaptive Differential Evolution 6 / 29

8 The Differential Evolution Algorithm The Differential Evolution Algorithm Price & Storn, 1995 (JGO:1997) NP.. population size (D-dimensional vectors) F.. mutation scale factor CR.. crossover parameter rand/1 strategy: v (G) i = x r (G) 1 + F ( x (G) r2 x (G) r3 ), r 1 r 2 r 3 i { u (G) v (G) i,j if rand(0, 1) CR or j = j rand, i,j = x (G) i,j otherwise, where i = 1, 2,..., NP and j = 1, 2,..., D. Dynamic Optimization using Self-Adaptive Differential Evolution 6 / 29

9 The Differential Evolution Algorithm The Differential Evolution Algorithm Price & Storn, 1995 (JGO:1997) NP.. population size (D-dimensional vectors) F.. mutation scale factor CR.. crossover parameter rand/1 strategy: v (G) i = x r (G) 1 + F ( x (G) r2 x (G) r3 ), r 1 r 2 r 3 i { u (G) v (G) i,j if rand(0, 1) CR or j = j rand, i,j = x (G) i,j otherwise, where i = 1, 2,..., NP and j = 1, 2,..., D. x, u better survives Dynamic Optimization using Self-Adaptive Differential Evolution 6 / 29

10 The Self-adaptive DE Algorithm The Self-adaptive DE Algorithm: jde algorithm F (G+1) i = CR (G+1) i = { Fl + rand 1 F u if rand 2 < τ 1, otherwise, { rand3 if rand 4 < τ 2, F (G) i otherwise. τ 1 = 0.1, τ 2 = 0.1, F l = 0.1, F u = 0.9 (fixed values) F [0.1, 1.0], CR [0, 1] CR (G) i [4] J. Brest et al. Self-Adapting Control Parameters in Differential Evolution: A Comparative Study on Numerical Benchmark Problems. IEEE TEVC, 10(6): , Dynamic Optimization using Self-Adaptive Differential Evolution 7 / 29

11 1 Introduction 2 Background The Differential Evolution Algorithm The Self-adaptive DE Algorithm 3 Algorithm for DOPs 4 Experimental Results 5 Conclusions Dynamic Optimization using Self-Adaptive Differential Evolution 8 / 29

12 Hybridized algorithm for solving Dynamic Optimization Problems (DOPs) multi-populations (random indexes r 1, r 2, and r 3 indicate vectors (individuals) that belong to same subpopulation as the trial vector x i ) self-adaptive control mechanism, F belongs to interval [0.36, 1] (F l = 0.36 suggested by D. Zaharie [22]) aging at individual level (an individual that stagnates in local optimum should be reinitialized) overlapping search between two subpopulations (distance of the best individuals of the subpopulations) reinitialization (when individual is close to local best) archive (currently best individual is added to archive after each change is detected) Dynamic Optimization using Self-Adaptive Differential Evolution 9 / 29

13 Our algorithm Multi-populations random indexes r 1, r 2, and r 3 indicate vectors (individuals) that belong to same subpopulation as the trial vector x i ) more populations without any information sharing (except overlapping search between two best individuals of two subpopulations) subpopulations search different regions diversity is important feature in DOPs Dynamic Optimization using Self-Adaptive Differential Evolution 10 / 29

14 F belongs to interval [0.36, 1] D. Zaharie [22] critical values for the control parameters of DE 2F 2 2/NP + CR/NP = 0... can be considers to be critical (this formula has an error in the paper ) assume CR = 0 and NP = 10 then critical value for F is (0.424 when NP = 5) we set F l = 0.36 in all experiments Dynamic Optimization using Self-Adaptive Differential Evolution 11 / 29

15 Our algorithm aging at individual level an individual that stagnates in local optimum should be reinitialized each individual has its own age-variable (incremented once per generation) three rules for aging (see Alg. 1): global best is not reinitialized when local best needs to be reinitialized, the whole subpopulation with some probability is reinitialized another individual is reinitialized with some probability Dynamic Optimization using Self-Adaptive Differential Evolution 12 / 29

16 Our algorithm Individual s improvement and aging when an improvement of individual occurs the age is set to some small value the new promising individual should stay in population for more generations the distance measure and fitness are used to make decision when individual s improvement is small or big (see Alg. 4) Dynamic Optimization using Self-Adaptive Differential Evolution 13 / 29

17 Our algorithm Archive algorithm starts with empty archive currently best individual is added to the archive, after each change is detected an individual is selected from archive only for the first subpopulation Dynamic Optimization using Self-Adaptive Differential Evolution 14 / 29

18 Parameter Settings F self-adaptive, CR self-adaptive, NP = 50, number of sub-populations: 5 (the size of each sub-populations was 10). Dynamic Optimization using Self-Adaptive Differential Evolution 15 / 29

19 1 Introduction 2 Background The Differential Evolution Algorithm The Self-adaptive DE Algorithm 3 Algorithm for DOPs 4 Experimental Results 5 Conclusions Dynamic Optimization using Self-Adaptive Differential Evolution 16 / 29

20 Results Table: CEC 09 Dynamic Optimization benchmark functions F 1 F 2 F 3 F 4 F 5 F 6 Rotation peak function Composition of Sphere s function Composition of Rastrigin s function Composition of Griewank s function Composition of Ackley s function Hybrid Composition function Dynamic Optimization using Self-Adaptive Differential Evolution 17 / 29

21 Results Table: Error Values Achieved for Problems F 1 Dimension(n) Peaks(m) Errors T 1 T 2 T 3 T 4 T 5 T Avg best Avg worst Avg mean STD Avg best Avg worst Avg mean STD T 7 (5-15) 10 Avg best 0 Avg worst Avg mean STD Avg best 0 Avg worst Avg mean STD Dynamic Optimization using Self-Adaptive Differential Evolution 18 / 29

22 Results Table: Error Values Achieved for Problems F 2 Dimension(n) Errors T 1 T 2 T 3 T 4 T 5 T 6 10 Avg best Avg worst Avg mean STD T 7 (5-15) Avg best 0 Avg worst Avg mean STD Dynamic Optimization using Self-Adaptive Differential Evolution 19 / 29

23 Results Table: Error Values Achieved for Problems F 3 Dimension(n) Errors T 1 T 2 T 3 T 4 T 5 T 6 10 Avg best e e e e-14 Avg worst Avg mean STD T 7 (5-15) Avg best 0 Avg worst Avg mean STD Dynamic Optimization using Self-Adaptive Differential Evolution 20 / 29

24 Results Table: Error Values Achieved for Problems F 4 Dimension(n) Errors T 1 T 2 T 3 T 4 T 5 T 6 10 Avg best Avg worst Avg mean STD T 7 (5-15) Avg best 0 Avg worst Avg mean STD Dynamic Optimization using Self-Adaptive Differential Evolution 21 / 29

25 Results Table: Error Values Achieved for Problems F 5 Dim.(n) Errors T 1 T 2 T 3 T 4 T 5 T 6 10 Avg best e e e e e e-14 Avg worst Avg mean STD T 7 (5-15) Avg best e-14 Avg worst Avg mean STD Dynamic Optimization using Self-Adaptive Differential Evolution 22 / 29

26 Results Table: Error Values Achieved for Problems F 6 Dimension(n) Errors T 1 T 2 T 3 T 4 T 5 T 6 10 Avg best Avg worst Avg mean STD T 7 (5-15) Avg best 0 Avg worst Avg mean STD Dynamic Optimization using Self-Adaptive Differential Evolution 23 / 29

27 Results Table: Algorithm Overall Performance F 1 (10) F 1 (50) F 2 F 3 F 4 F 5 F 6 T T T T T T T Mark Performance (sumed the mark obtained for each case and multiplied by 100): Dynamic Optimization using Self-Adaptive Differential Evolution 24 / 29

28 Convergence graphs F 1 F 4 Dynamic Optimization using Self-Adaptive Differential Evolution 25 / 29

29 Convergence graphs F 5, and F 6 Dynamic Optimization using Self-Adaptive Differential Evolution 26 / 29

30 Discussion our algorithm performs very well on small step (T 1 ) and chaotic (T 4 ) change types for F 1 F 4 F 5 : it obtained good results over all changed types F 6 : it obtained very well results over all changed types F 3 is the most difficult one among all test problems Dynamic Optimization using Self-Adaptive Differential Evolution 27 / 29

31 Conclusions jde algorithm with multi-populations and aging mechanism was evaluated on CEC 09 test problems special session on dynamic optimization problems. Overall performance: 69.7 Future plans: to apply additional co-operation among sub-populations to use sub-populations of different sizes to improve the usage of the archive (here, a simple variant of the archive is used) Dynamic Optimization using Self-Adaptive Differential Evolution 28 / 29

32 Thank You Questions? Dynamic Optimization using Self-Adaptive Differential Evolution 29 / 29