Linearized Mathematical Programming approach - ULg
Transcription
Linearized Mathematical Programming approach - ULg
RECONCILIATION OF MATHEMATICAL PROGRAMMING AND OPTIMALITY CRITERIA APPROACHES IN STRUCTURAL OPTIMIZATION Part II: Linearized MP approach Pierre DUYSINX Patricia TOSSINGS LTAS – Automotive Engineering Academic year 2015-2016 1 LAY-OUT Generalized optimality criteria Linearized mathematical programming methods Unified approach to structural optimization Examples 2 INTRODUCTION 3 INTRODUCTION Optimality criteria techniques (OC) – Highly specific – Intuitive techniques, simple – Convergence to a design that is not necessarily optimal (KKT conditions) – Difficulties in identifying the set of active constraints – Convergence instabilities – Small number of reanalyses, independent of the number of design variables Résumé – Low cost – But uncertainty convergence 4 INTRODUCTION Pure Mathematical Programming methods – Very general – Rigorous methods, quite elaborated – Convergence to a local minimum – Stable and monotonic convergence – Large number of reanalyses, growing with the number of design variables Résumé – Rigorous framework & guaranteed convergence – High cost (Growing with the size of the problem) 5 INTRODUCTION Generalized optimality criteria – Dual scheme to solve the set of active constraints – Inherent mechanism of selection of the set of active constraints Mixed methods – Modification of pure MP methods to reduce the number of function evaluations – Projection scheme – Linearization of constraints – Restoring algorithm – Controlling the convergence of the optimization process 6 INTRODUCTION Relation between optimality criteria (OC) and Mathematical Programming (MP) approaches – Generalized optimality criteria = Math Programming linearization methods – First order explicit approximation – Constraint gradients – Stress ratioïng – Continuous transition between strict primal MP methods and pure OC methods 7 INTRODUCTION UNIFIED OPTIMIZATION APPROACH – First order approximation concept approach – Sequence of explicit subproblems with linear or convex approximations of constraints – Efficient solution of the subproblems using primal and dual mathematical programming algorithms Dual schemes generalized optimality criteria Primal scheme mixed method 8 MATHEMATICAL PROGRAMMING METHODS WITH LINEARIZATION Three actions: – Use of reciprocal variables – Restoration phase using scaling of variables – Linearization = high quality approximations Linearization approach – Primal and mixed linearization methods 9 Mathematical programming approach Primal projection method: gradient projection method with NL constraints – Steadily improving (decreasing) objective function – Always feasible designs Several structural (FE) reanalyses / iterations Large number of iterations : Growing with the number of design variables PROHIBITIVE COST FOR LARGE SCALE PROBLEMS 10 Mathematical programming approach 1. Feasible boundary point x(0). 2. Minimization phase: 1. Search direction s= gradient projected onto the restriction tangent planes 2. Step length is determined by line search 1. Requires serveral reanalyses! 2. x(1) = minimum along s 3. 4. Restoration phase 1. Non linear constraints: x(1) is generally non feasible 2. Find x(2) back to the boundary of the feasible domain 3. Iterative process: it also requires several reanalyses Feasible boundary point x(2): new minimization phase go to step 2 11 Mathematical programming approach Three actions to be able to apply MP to structural and multidisciplinary optimization 1/ REDUCE THE NON LINEAR CHARACTER OF THE CONSTRAINTS Change of variables: switch to the reciprocal design space New constraints very shallow (assumption: weak structural redundancy) Much larger steps can be taken without seriously violating the constraints 12 Mathematical programming approach 2/ RESTORATION PHASE : Use scaling (No internal force redistribution during scaling) Scaling factor determined by the most violated constraint Only one reanalysis per analysis / restoration 13 Mathematical programming approach 3/ APPROXIMATE THE LINE SEARCH DURING THE MINIMIZATION PHASE Instead of using an exact line search that costs several FE analyses, consider that linearized behavior constraint in the reciprocal constraints are very good explicit approximations Skip reanalyses: accept solutions if feasible according to approximations CONCLUSION: Only one single reanalysis per iteration 14 Projection and restoration with mixed MP 15 Mathematical programming approach 3 actions One single FE reanalysis per iteration BUT Number of iteration is still growing with the number of design variables Proposed approach – Sequence of several search directions evaluated without reanalyses – Periodic update of the linearized constraints (after k search directions) – Linearized approximation in the reciprocal design variables are high quality explicit approximations of the responses constraints 16 Mathematical programming approach Linearized problem – New minimization phase = solving partly this problem by gradient projected method for linear constraints – Perform k iterations then reanitialization (update) 17 Mathematical programming approach Number of solution steps of the subproblem solution is a convergence control parameter: – PRIMAL APPROACH : k small: Sequence of steadily improved feasible design Pure projection method k=1 Monotonic convergence – LINEARIZATION APPROACH : k large: Sometimes increase of the weight after rescaling because of the constraint violations. More or less complete solution of the subproblem Pure linearization method (k ∞) Fast but risk of unstable convergence 18 Mathematical programming approach Ten-bar truss – Mixed methods Stress and displacement constraints 19 RELATIONS BETWEEN OC AND MP Generalized OC = MP linearization method Approximation concept approach Constraint gradients 20 Generalized OC = MP linearization method GOC : sequence of explicit subproblems where real constraints are approximated by – cij = virtual energy densities Mixed method: sequence of linearized problems with – cij = derivatives of the responses functions with respect to the reciprocal variables 21 Generalized OC = MP linearization method Reminder: Equality of the derivatives wrt to reciprocal variables and of the virtual strain energy For OC: the strain energy density For the mixed method 22 Generalized OC = MP linearization method For the mixed method – The change of variable to the reciprocal variable So it comes 23 Generalized OC = MP linearization method Unified approach: – Sequence of explicit subproblems obtained by linearizing the behavior constraints with respect to the reciprocal variables Later: linearizing any behavior constraints with respect to the reciprocal variables! – Independence wrt the number of design variables! Solution of the explicit subproblems – Dual solution scheme: generalization of conventional OC techniques (GOC) – Primal solution scheme: mixed method: gradual transition between pure MP and OC approaches 24 Generalized OC = MP linearization method Moderate structural redundancy: – shallow constraints – Non linearity in reciprocal space is weak OC OK! Strong structural redundancy: – highly non linear constraint mixed approach can help! Convergence of the OC and mixed methods 25 Approximation concepts (MP methods) Brief history of MP approaches – Primal methods: Feasible direction: Schmidt, Gellaty, Tocher, Vanderplats… Gradient projection: Brown and Ang – Barrier functions: Moe, Kavlie – Linearization methods: Moses, Pope Pedersen Prohibitive cost Convergence to a vertex Instabilities 26 Approximation concepts (MP methods) Use of reciprocal variables – Pure linearization techniques: Reinschmidt et al. – Pure linearization using inscribed hyper spheres: Schmidt and Farshi – Objective function not linearized: Schmidt and Miura Approximation concept approach – Basically the same approach as mixed methods – Primal philosophy: partial solution of the linearized problem Feasible directions (CONMIN) Extended barrier function (NEWSUMT) – Dual methods (DUAL2, DUAL1, CONLIN, MMA…) MP have evolved towards OC techniques 27 Approximation concepts (MP methods) Symmetry 2 design variables Stress constraints only Optimum: only one active constraint (not FSD) 28 Approximation concepts (MP methods) Three bar truss problem: trajectories in the reciprocal design space Solution by ACCESS 3 with NEWSUMT and DUAL2 Strictly feasible approach Follow the constraint Newsumt (0.5 x 1): -0.5 response factor decrease ratio -1: number of unconstrained minimization 29 Approximation concepts (MP methods) Three bar truss problem: convergence of the weight (objective function) Convergence acceleration as the explicit subproblems are solved more deeply 30 NUMERICAL APPLICATIONS Comparison of 1st and 2nd order algorithms 72 bar truss 25 bar truss 200 bar truss 63 bar truss Composite box beam I beam Delta wing Aircraft spoiler 31 Optimization algorithms Sequence of linearly constrained subproblems General purpose optimizers – CONMIN: feasible directions – NEWSUMT: barrier function – PRIMAL1: gradient projection 32 Optimization algorithms Because of the simple algebraic structure: – Explicit – Strictly convex – Separable Second order primal and dual algorithms – Primal second order algo: partial solution of the optimization subproblem is possible (mixed method) – Dual algo: complete solution of the subproblem Equivalent to Generalized OC techniques – Based on Newton’s method 33 Test problem Primal vs Dual solutions 34 72 bar truss problem (four level tower) Stress and minimum size constraints Displacement limits at nodes 1, 2, 3, 4 (X and Y directions) Symmetry 16 design variables (using dv linking) Two load cases 35 72 bar truss problem (four level tower) Stress constraints not very critical treated by stress ratioing (not expensive) 36 63 bar truss: Wing Carry through box Two load cases Upper limit on the relative displacement in the X direction (torsion rotation) Stress and minimum size constraint No linking : 63 independent design variables 37 63 bar truss: Wing Carry through box 38 63 bar truss: Wing Carry through box 39 63 bar truss: Wing Carry through box Much lower computation al cost Faster convergence but CPU increase Slow and unstable convergence 40 I- beam 41 I- beam 42 Delta Wing problem Delta wing structure Fiber reinforced skins (Carbone epoxy) – 0°/45°/90°/-45° – Symmetric and balanced laminates Metallic webs Single load case (pressure) Temperature change conditions 43 Delta Wing problem Thickness of each composite ply i.e. 252 orthotropic membranes for skins Thickness of aluminum webs: 70 symmetric shear panels Design variable linking 60 design variables Constraints: deflection of wing tip nodes Max strain criteria in composite panels Lower bound on eigenfrequency with fixed mass (fuel) 44 Delta Wing problem 45 Delta Wing problem 46 Weight minimization of an aircraft spoiler Light aluminum alloy Front spar and secondary spar joined by 12 ribs And covered by 2 skins 2 load cases (pressure) Stress constraints (Fleury, 1976) Landing: trailing edge must remain straight within a tolerance e< 0.5 mm Difficult displacement constraint for any 47 nodes on trailing edge Weight minimization of an aircraft spoiler 627 design variables OC techniques failed Mixed methods and Sequential programming have to be used (Fleury, 1976) 48 Weight minimization of an aircraft spoiler Two simplified FE models 49 Weight minimization of an aircraft spoiler (Fleury, 1976) 50 Weight minimization of an aircraft spoiler 51 Weight minimization of an aircraft spoiler 52 Weight minimization of an aircraft spoiler 53 Weight minimization of an aircraft spoiler 54 (Fleury, 1976) Weight minimization of an aircraft spoiler 55 (Fleury, 1976) Weight minimization of an aircraft spoiler 56 (Fleury, 1976) CONCLUSION 57 CONCLUSION Sequence of explicit subproblems – First Order Approximations Primal / dual solution schemes DUAL Generalized of OC Computationally economical but convergence instability Discrete design variable possible Reliable computer implementation Dual bound = monitoring PRIMAL Mixed method (OC/PM) Control over convergence at a higher cost Other objective functions non separable explicit functions Sophisticated algorithms 58