Constrained Optimization by the e Constrained Differential Evolution with an Archive and Gradient-Based Mutation Tetsuyuki TAKAHAMA (Hiroshima City University) Setsuko SAKAI (Hiroshima Shudo University) Outline Constrained optimization problems The e constrained method Constraint violation and e-level comparisons The e constrained differential evolution (eDEag) differential evolution (DE) with an archive gradient-based mutation control of the e-level Experimental results Conclusions 2010/07/21 T.Takahama and S.Sakai in CEC2010 2 Constrained Optimization Problems objective function f , decision variables xi inequality constraints gj, equality constraints hj lower bound li, upper bound ui 2010/07/21 T.Takahama and S.Sakai in CEC2010 3 e constrained method Algorithm transformation method algorithm for unconstrained optimization → algorithm for constrained optimization e-level comparison comparison between pairs of objective value and constraint violation replacing ordinary comparisons to e-level comparisons in unconstrained optimization algorithm by 2010/07/21 T.Takahama and S.Sakai in CEC2010 4 Constraint Violation Constraint Violation f (x) f ( x ) 0 , if x is infeasible f ( x ) 0 , if x is feasible max f ( x ) max{ max { 0 , g j ( x )}, max | h j ( x ) |} j sum f (x) || max{ j 2010/07/21 j 0 , g j ( x )} || || h j ( x ) || p p j T.Takahama and S.Sakai in CEC2010 5 e-level comparison Function value and constraint violation(f ,f) 2010/07/21 precedes constraint violation usually precedes function value if violation is small T.Takahama and S.Sakai in CEC2010 6 Definition of e constrained method ∥ Constrained problems can be solved by replacing ordinary comparisons with e level comparisons in unconstrained optimization algorithm 2010/07/21 <→<e, → e T.Takahama and S.Sakai in CEC2010 7 Differential Evolution (DE) simple operation avoiding step size control crossover parent population base vector - F difference vector (CR) + trial vector trial vector (child) will survive if the child is better robust to non-convex, multi-modal problems 2010/07/21 T.Takahama and S.Sakai in CEC2010 8 eDEa: eDE with an archive (1) A small population and a large archive are adopted N Generate M initial individuals Small population is good for search efficiency but is bad for diversity A={ xk | k=1,2,...,M } (M=100n) Select top N individuals from A as an initial population 2010/07/21 P P={ xi | i=1,2,...,N } (N=4n) A=A-P A T.Takahama and S.Sakai in CEC2010 M-N 9 eDE with an archive (2) DE/rand/1/exp operation correction of Fig.2 mutant vector: and are selected from P is selected from P A w.p. 0.95 or P w.p. 0.05 exponential crossover Uniform convergence of individuals When a parent generates a child and the child is not better than the parent, the parent can generate another child 2010/07/21 T.Takahama and S.Sakai in CEC2010 10 eDE with an archive (3) Direct replacement for efficiency Continuous generation model If the child is better than the parent, the parent is directly replaced by the child (f(xtrial),f(xtrial)) <e (f(xi), f(xi)) Perturb scaling factor F in small probability to escape from local minima F is a fixed value (0.5) w.p 0.95 F=1+|C(0,0.05)| truncated to 1.1 w.p. 0.05 2010/07/21 T.Takahama and S.Sakai in CEC2010 11 Gradient-based mutation (1) adopts the gradient of constraints to reach feasible region Constraint vector and constraint violation vector T C( x ) (g 1 ( x ) , , g q ( x ) ,h q 1 ( x ) , h m ( x )) C( x ) ( g 1 ( x ) , , g q ( x ) ,h q 1 ( x ) , h m ( x )) T g j ( x ) max{ 0 , g j ( x )} Gradient of constraint vector C( x ) x C( x ) if satisfied, 2010/07/21 x x will be feasible T.Takahama and S.Sakai in CEC2010 12 Gradient-based mutation (2) inverse C( x ) 1 cannot be defined generally Moore-Penrose inverse (pseudoinverse) C( x ) approximate or best (LSE) solution x ' x C( x ) C( x ) Modifications Numerical gradient (costs n+1 FEs) Mutation is applied only in every n generations Skipped w.p. 0.5, if num. of violated constraints is one 2010/07/21 T.Takahama and S.Sakai in CEC2010 13 Control of e-level Small feasible region and e-level small feasible region f=0 f≦ e(Tf) f≦ e(0) 2010/07/21 T.Takahama and S.Sakai in CEC2010 14 Control scheme of e-level e-level should converge to 0 gradually f ( x ) cp t e (t ) e ( 0 ) 1 Tc 0 e(t) e0 (t 0 ) ( 0 t Tc ) (t Tc ) x is the top - th individual 0 2010/07/21 Tc Tmax in f t T.Takahama and S.Sakai in CEC2010 15 control of cp instead of specifying cp, specify e-level at Tl , To search better objective value generation enlarge 2010/07/21 from Tl to Tc e-level and scaling factor F T.Takahama and S.Sakai in CEC2010 16 Effectiveness of e constrained method The e level comparison does not need objective values if one of the constraint violations is larger than e-level Lazy evaluation objective function is evaluated only when needed evaluation of objective function can be often omitted when feasible region is small 2010/07/21 T.Takahama and S.Sakai in CEC2010 17 Conditions of experiments 18 constrained problems, 25 trials per a problem eDEag/rand/1/exp Max. FEs: 20,000n M=100n, N=4n, F=0.5, CR=0.9 e level control: =0.9, Tc=1,000, Tl=0.95Tc Gradient-based mutation 2010/07/21 mutation rate: Pg=0.1, max. iterations: Rg=3 applied only in every n generations T.Takahama and S.Sakai in CEC2010 18 Summary of Results Feasible and stable solutions in all runs 10D: C01-C07, C09, C10, C12-C14, C18 (13) 30D: C01, C02, C05-C08, C10, C13-C16 (11) Feasible solutions in all runs 10D: C08, C11, C15, C16, C17 (5) 30D: C03, C04, C09, C11, C17, C18 (6) Often infeasible solutions 30D: 2010/07/21 C12 (1) T.Takahama and S.Sakai in CEC2010 19 10D (C01-C06) 2010/07/21 T.Takahama and S.Sakai in CEC2010 20 10D (C07-C012) 2010/07/21 T.Takahama and S.Sakai in CEC2010 21 10D (C13-C018) 2010/07/21 T.Takahama and S.Sakai in CEC2010 22 30D (C01-C06) 2010/07/21 T.Takahama and S.Sakai in CEC2010 23 30D (C07-C12) 2010/07/21 T.Takahama and S.Sakai in CEC2010 24 30D (C13-C18) 2010/07/21 T.Takahama and S.Sakai in CEC2010 25 Computational Complexity T1: Time (seconds) of 10,000 function evaluations for a problem on average T2: Time (seconds) of 10,000 function evaluation with algorithm for a problem 2010/07/21 T.Takahama and S.Sakai in CEC2010 26 Conclusions eDE with a large archive and gradient-based mutation can find feasible solutions in all run and all problems except for C12 of 30D can often omit evaluation of objective values and find solutions efficiently and very fast 2010/07/21 T.Takahama and S.Sakai in CEC2010 27 Future works To find better objective values dynamic control of e level e level according to the number of feasible points changing mechanism for maintaining diversity subpopulations or species to search various regions adaptive 2010/07/21 control of F and CR T.Takahama and S.Sakai in CEC2010 28 Thank you for your kind attention 2010/07/21 T.Takahama and S.Sakai in CEC2010 29 10D problems 2010/07/21 T.Takahama and S.Sakai in CEC2010 30 10D problems 2010/07/21 T.Takahama and S.Sakai in CEC2010 31 30D problems 2010/07/21 T.Takahama and S.Sakai in CEC2010 32 30D problems 2010/07/21 T.Takahama and S.Sakai in CEC2010 33 Moore-Penrose inverse A : pseudo inverse singlar va lue decomposit A UV T : inverting non - zero elements on the diagonal 2010/07/21 ion T A V U of A of T.Takahama and S.Sakai in CEC2010 34