Modified Differential Evolution for Nonlinear Optimization Problems with Simple Bounds

Transcription

1 Modified Differential Evolution for Nonlinear Optimization Problems with Simple Bounds Md. Abul Kalam Azad a,, Edite M.G.P. Fernandes b a Assistant Researcher, b Professor Md. Abul Kalam Azad Algoritmi R&D Centre akazad@dps.uminho.pt URL: Department of Production and Systems School of Engineering, University of Minho, Portugal Presented By: May 2, 2009 M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

2 Outline of the Presentation 1 Introduction 2 Motivation 3 Differential Evolution 4 Modified Differential Evolution 5 Algorithm of Proposed Modified Differential Evolution 6 Experimental Results 7 Conclusions M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

3 Introduction CMS2009, May 1-3, 2009 Problems involving global optimization over continuous spaces are ubiquitous throughout the scientific community. The task is to optimize certain properties of a system by pertinently choosing its variables. The task of global optimization is to find a point where the objective function obtains its smallest value. Nonlinear optimization problems with simple bounds where min f(x) subject to x Ω, f : R n R with Ω = {x R n : lb j x j ub j, j = 1,...,n}, and lb, ub R n. (1) M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

4 Introduction CMS2009, May 1-3, 2009 There exist many solution methods to solve (1). If the objective function is not differentiable or has no information about derivatives: Deterministic methods that guarantee to find a global optimum with a required accuracy: DIRECT, MCS, Pattern Search, Simplex Search, etc. Stochastic methods that find the global minimum only with high probability: GA, SA, PSO, DE, EM, etc. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

5 Motivation Differential Evolution (DE) is proposed by Storn and Price in DE is a population based heuristic approach. DE has only three parameters. DE has been shown to be very efficient. DE is applicable to derivative free nonlinear optimization problems. Proposed Method A modified differential evolution (mde) introducing self-adaptive parameters and the inversion operator for nonlinear optimization problems with simple bounds. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

6 Differential Evolution (DE) DE creates new candidate solutions by combining points of the same population. A candidate replaces a current solution only if it has better objective value. DE has three parameters: 1 Amplification factor of the difference vector F. 2 Crossover control parameter CR. 3 Population size NP. DE s operators 1 Mutation 2 Crossover 3 Selection M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

7 Outline of Differential Evolution Target Vector n- problem dimension, NP- population size. Target point is defined by x p,t = (x p1,t, x p2,t,...,x pn,t ) where t is the index of generation and p = 1, 2,...,NP. The target point at t = 1 is generated randomly with uniform distribution as x p,1 = lb + r (ub lb) (2) r U[0, 1]. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

8 Outline of Differential Evolution Mutation DE creates new points (mutant points) by adding the weighted difference between two points to a third point in a population. v p,t+1 = x r1,t +F(x r2,t x r3,t) (3) }{{}}{{} base differential Integer random numbers r 1, r 2, r 3 U{1, 2,...,NP} and r 1 r 2 r 3 p. F [0, 2] constant parameter which controls the amplification of the differential variation (x r2,t x r3,t). NP must be greater or equal to 4 M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

9 Outline of Differential Evolution Crossover [Trial Point] The mutant point s components are then mixed with the target point s components to yield the so-called trial point u p,t+1. { vpj,t+1 if (r u pj,t+1 = j CR) or j = z p, x pj,t if (r j > CR) and j z p j = 1, 2,...,n. (4) Random r j U[0, 1] which performs the mixing of jth component of points and CR [0, 1] is constant parameter. Random integer z p U{1, 2,...,n} which ensures that u p,t+1 gets at least one component from v p,t+1. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

10 Outline of Differential Evolution Bounds Check After crossover the bounds of each component must be checked. lb j if u pj,t+1 < lb j u pj,t+1 = ub j if u pj,t+1 > ub j otherwise u pj,t+1 (5) Selection The trial point u p,t+1 is compared to the target point x p,t to decide whether or not it should become a member of generation t + 1 as { up,t+1 if f(u x p,t+1 = p,t+1 ) f(x p,t ) otherwise. x p,t (6) M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

11 Modified Differential Evolution According to Storn and Price, DE is sensitive to the choice of three control parameters. They suggested (i) F [0.5, 1]; (ii) CR [0.8, 1]; and (iii) NP = 10 n. Modification by Brest et al. Proposed self adaptive control parameters for F and CR. Point dependent parameters (F p,1, CR p,1 ), p = 1,...,NP. New control parameters for next generation F p,t+1 and CR p,t+1 are calculated as M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

12 Modification CMS2009, May 1-3, 2009 Modification by Brest et al. { Fl + λ F p,t+1 = 1 F u, if λ 2 < τ 1 F p,t, otherwise { λ3, if λ CR p,t+1 = 4 < τ 2 CR p,t, otherwise. (7) (8) Random λ U[0, 1] and τ 1 = τ 2 = 0.1 represent probabilities to adjust parameters F p and CR p, respectively. F l = 0.1 and F u = 1.0. So F p,t+1 [0.1, 1.0] and CR p,t+1 [0, 1]. F p,t+1 and CR p,t+1 are obtained before the mutation is performed. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

13 Modification CMS2009, May 1-3, 2009 Modification by Kaelo et al. Kaelo et al. proposed alternative technique for mutation. After choosing three points the best is selected for base point, Remaining two points are used as differential variation: v p,t+1 = x 1,t +F(x 2,t x 3,t ) }{{}}{{} base differential x 1,t = arg min{f(x r1,t), f(x r2,t), f(x r3,t)}. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

14 Our Modified Differential Evolution (mde) Our mde has previous modifications proposed by Kaelo et al. and Brest et al. We made small modification proposed by Kaelo et al. After every B generations we used best point found so far as the base point and two randomly chosen points in differential variation: v p,t+1 = x best,t +F(x r1,t x r2,t) }{{}}{{} base differential Inversion operator Our mde has inversion operator with some inversion probability (p inv [0, 1]) to points after crossover. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

15 Our Modified Differential Evolution (mde) Illustrative example of inversion: h k u p,t = u p1,t u p2,t u p3,t u p4,t u p5,t u p6,t u p7,t u p8,t h k u p,t = u p1,t u p2,t u p6,t u p5,t u p4,t u p3,t u p7,t u p8,t Termination condition t- current generation G max - maximum generations. The terminition condition is ((t > G max ) OR ( f max, t f min, t ɛ (= 10 6 ))). f max, t - max. function value f min, t - min. function value M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

16 Algorithm of Proposed Modified Differential Evolution 1 Set the values of parameters. 2 Set t = 1. Initialize the population x p, 1. 3 Calculate f max, 1 and f min, 1 and set f best = f min, 1 and x best = x min, 1. 4 If termination condition is met stop. Otherwise set t = t Compute mutant point v p, t by using mde. 6 Perform crossover by using mde to make point u p, t. 7 Perform inversion by using mde to make trial point u p, t. 8 Check the domains of the trial point. 9 Perform selection. If f p, t f p, t 1 then set x p, t = u p, t. Otherwise set f p, t = f p, t 1 and x p, t = x p, t Calculate f max, t and f min, t and set f best = f min, t and x best = x min, t. 11 Go to step 4. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

17 Experimental Results We coded all the variants of DE in C with AMPL interface. The name of the solvers are: DE Original mde[1] DE Kaelo mde[2] (with inversion) DE Brest We tested all the solvers on a set of 64 nonlinear optimization problems with simple bounds. We used same parameters for all solvers. We used same termination condition for all solvers. We run all solvers 30 times for each problem. After 30 runs we reported f best, f w, f avg, standard deviation of f, no. of function evaluations and no. of generations. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

18 Experimental Results Example Test Problems Ackley s Problem min x f(x) n n 20 exp 0.02 n 1 xj 2 exp n 1 cos(2πx j ) exp(1) j=1 j=1 subject to 30.0 x j 30.0, j = 1, 2,..., n. Griewank Problem min x f(x) n j=1 x 2 j ( ) n x j cos j j=1 subject to x j 600.0, j = 1, 2,...,n. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

19 Experimental Results Shekel 5 Problem 5 1 min f(x) x n i=1 j=1 (x j a ij ) 2 + c i subject to 0.0 x j 10.0, j = 1, 2,...,n. For the above 3 test problems for all solvers: Used same parameters. Used same termination condition. Plotted profile of objective function value at different generation. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

20 Experimental Results Ackley s Problem, n = 10, G max = 1000, f opt = 0.0 Profile of objective function value of ack after single run 2.5 DE_Original DE_Kaelo DE_Brest mde[1] mde[2] 2.0 Objective function value No. of generations M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

21 Experimental Results Griewank Problem, n = 10, G max = 1000, f opt = Profile of objective function value of gw after single run DE_Original DE_Kaelo DE_Brest mde[1] mde[2] Objective function value No. of generations M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

22 Experimental Results Shekel 5 Problem, n = 4, G max = 250, f opt = Profile of objective function value of s3 after single run DE_Original DE_Kaelo DE_Brest mde[1] mde[2] Objective function value No. of generations M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

23 Experimental Results Neumaier 3 Problem min x f(x) = n (x j 1) 2 j=1 n x j x j 1 j=2 subject to n 2 x j n 2, j = 1, 2,...,n. For above test problem for all solvers: Used n = 10, 20, 30, 40. Used same parameters. Used same termination condition. Run 30 times. Reported f best at every G max /50 generations and made average. Plotted profile of mean of best objective function value at every G max /50 generations. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

24 Experimental Results Neumaier 3 Problem, n = 10, G max = 500, f opt = Mean of best objective function value Profile of mean of best objective function value of nf3_10 after 30 runs 4500 DE_Original DE_Kaelo 4000 DE_Brest mde[1] 3500 mde[2] Mean of best objective function value Profile of mean of best objective function value of nf3_10 after 30 runs 3000 DE_Original DE_Kaelo DE_Brest 2500 mde[1] mde[2] No. of generations (a) No. of generations (b) M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

25 Experimental Results Neumaier 3 Problem, n = 20, G max = 1500, f opt = Mean of best objective function value 16 x Profile of mean of best objective function value of nf3_20 after 30 runs 104 DE_Original DE_Kaelo 14 DE_Brest mde[1] mde[2] No. of generations (a) Mean of best objective function value x 10 4 Profile of mean of best objective function value of nf3_20 after 30 runs DE_Original DE_Kaelo DE_Brest mde[1] mde[2] No. of generations (b) M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

26 Experimental Results Neumaier 3 Problem, n = 30, G max = 3000, f opt = Mean of best objective function value 12 x Profile of mean of best objective function value of nf3_30 after 30 runs 105 DE_Original DE_Kaelo DE_Brest 10 mde[1] mde[2] Mean of best objective function value 6 x 105 Profile of mean of best objective function value of nf3_30 after 30 runs DE_Original DE_Kaelo DE_Brest 5 mde[1] mde[2] No. of generations (a) No. of generations (b) M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

27 Experimental Results Neumaier 3 Problem, n = 40, G max = 4000, f opt = Mean of best objective function value x 10 6 Profile of mean of best objective function value of nf3_40 after 30 runs DE_Original 6 DE_Kaelo DE_Brest mde[1] mde[2] Mean of best objective function value 3.5 x Profile of mean of best objective function value of nf3_40 after 30 runs 106 DE_Original DE_Kaelo DE_Brest 3 mde[1] mde[2] No. of generations (a) No. of generations (b) M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

28 Performance Profile CMS2009, May 1-3, 2009 Compared all solvers based on performance profile proposed by Dolan and Moré. P- set of all problems and S- set of all solvers. Performance metric found by solver s S on problem p P is m (p,s) = f (p,s) f opt f w f opt. f (p,s) - average/best function value after 30 runs. f opt - optimum value of problem p. f w - worst function value after 30 runs (for all solvers). m (p,s) = 0 if f (p,s) f opt 1 if f (p,s) = f w f (p,s) f opt f w f opt otherwise M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35 m f opt f w f

29 Performance Profile CMS2009, May 1-3, 2009 Since min{m (p,s) : s S} can be 0, the performance ratios are r (p,s) = 1 + m (p,s) min{m (p,s) : s S}, if min{m (p,s) : s S} < ɛ m (p,s) min{m (p,s) : s S}, otherwise for p P, s S and ɛ = The overall assessment of performance of a particular solver s ρ s (τ) = n P τ n P. n Pτ - number of problems in P with r (p,s) τ. n P - total number of problems in P. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

30 Performance Profile ρ s (τ) is the probability (for solver s S) that the performance ratio r (p,s) is within a factor τ R of the best possible ratio. The function ρ s is the cumulative distribution function for the performance ratio. The value of ρ s (1) gives the probability that the solver s will win over the others in the set. However, for large values of τ, the ρ s (τ) measures the solver robustness. The solver with largest ρ s (τ) is the one that solves more problems in the set P. We plotted performance profile based on f avg and f best after 30 runs for all problems with all solvers. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

31 Performance Profile CMS2009, May 1-3, 2009 performance profile based on f avg after 30 runs 1 Performance profile of f_avg after 30 runs ρ(τ) DE_Original 0.5 DE_Kaelo DE_Brest mde[1] mde[2] τ M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

32 Performance Profile CMS2009, May 1-3, 2009 performance profile based on f best after 30 runs 1.0 Performance profile of f_best after 30 runs 0.9 ρ(τ) DE_Original DE_Kaelo DE_Brest mde[1] mde[2] τ M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

33 Conclusions A modified differential evolution (mde) for nonlinear optimization problems with simple bounds is presented. A compartive study based on performance profile is presented. It is shown that our mde outperformed other variants of DE. Our mde wins over other DE based on robustness. For particular problem our mde also wins over other variants of DE. Future Study Now we are trying to implement our mde to general constrained nonlinear optimization problems and will consider mixed-integer problems in the future. M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

34 References CMS2009, May 1-3, 2009 [Ali2004] M.M. Ali and A. Törn, Population set based global optimization algorithms: Some modifications and numerical studies, Computer and Operration Research, vol. 31, no. 10, pp , [Ali2005] M.M. Ali, C. Khompatraporn and Z.B. Zabinsky, A numerical evaluation of several stochastic algorithms on selected continuous global optimization test problems, Journal of Global Optimization, vol. 31, pp , [Boender1982] C.G.E. Boender, A.H.G. Rinnoy Kan, L. Stougie and G.T. Timmer, A Stochastic Method for Global Optimization, Mathematical Programming, vol. 22, pp , [Brest2006] J. Brest, S. Greiner, B. Bošković, M. Mernik and V. Žumer, Self-adapting control parameters in differential evolution: a comparative study on numerical benchmark problems, IEEE Transactions on Evolutionary Computation, vol. 10. no. 6, pp , [Dolan2002] E.D. Dolan and J.J. Moré, Benchmarking optimization software with performance profiles, Mathematical Programming, series A 91, pp , [Kaelo2006] P. Kaelo and M.M. Ali, A numerical study of some modified differencial evolution algorithms, EJOR, vol. 169, pp , [Kim2007] H.-K. Kim, J.-K. Chong, K.-Y. Park and D.A. Lowther, Differential evolution strategy for constrained global optimization and application to practical engineering problems, IEEE Transactions on Magnetics, vol. 43, no. 4, pp , [Price1997] K. Price and R. Storn, Differential evolution - a simple evolution strategy for fast optimization, Dr. Dobb s Journal, vol. 22, no. 4, pp and 78, [Storn1997] R. Storn and K. Price, Differential evolution - a simple and efficient heuristic for global optimization over continuous spaces, Journal of Global Optimization, vol. 11, pp , M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35

35 Supported by under the grant C2007-UMINHO-ALGORITMI-04. and Thank You Very Much M. A. K. Azad (Algoritmi R&D Centre) University of Minho, Portugal May 2, / 35