FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING
Transcript of FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING
![Page 1: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/1.jpg)
FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING
Pierre DUYSINX
LTAS – Automotive Engineering
Academic year 2020-2021
1
![Page 2: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/2.jpg)
LAY-OUT
Optimality Criteria
– Fully stressed design
– Optimality criteria with one displacement constraint
Interpretation of OC as first order approximations
FSD zero order approximation
Berke’s approximation = 1st order approximation
2
![Page 3: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/3.jpg)
LAY-OUT
Optimality Criteria
– Optimality criteria with multiple displacement and stress constraints
Generalized optimality criteria
Dual maximization to solve OC criteria with multiple constraints
Unified approach to structural optimization
3
![Page 4: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/4.jpg)
4
OPTIMALITY CRITERIA :STRESS AND DISPLACEMENT
CONSTRAINTS
![Page 5: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/5.jpg)
5
Stress and displacement constraints
Combination of the two previous O.C.
– Stress constraints
– Displacement constraints
Set of active constraints is assumed to be known
– ň active design variables
– m active displacement constraints
Passive design variables: side constraints or determined by the stress constraints
![Page 6: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/6.jpg)
6
Stress and displacement constraints
Active design variables: ruled by displacement constraints
– m constraints
➔ m virtual load cases qj
➔ m Lagrange multipliers j
– Explicit approximation using virtual work
Lagrange function
![Page 7: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/7.jpg)
7
Stress and displacement constraints
Stationary conditions
After some algebra, the stationary conditions can be casted under the following form
If cij>0
![Page 8: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/8.jpg)
8
Stress and displacement constraints
For statically determinate case: OC are exact (because xi and cij
are constant)
➔ optimum in one analysis
For statically indeterminate case: OC are approximate
➔Iterative use of the redesign formulae
➔ Active variables
➔ Passive variables
![Page 9: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/9.jpg)
9
Stress and displacement constraints
Lagrange multipliers j????
– Such that the active displacement constraints are satisfied as equality
– Closed form solution only if m=1
– Otherwise numerical schemes
Envelop method (intuitive extension from case m=1)
Intuitive formula
Newton Raphson applied to solve the set of nonlinear equations
![Page 10: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/10.jpg)
10
Newton-Raphson iteration (Taig & Kerr, 1973)
Solve the system of nonlinear equations using a Newton-Raphson method
First set of equations enables to eliminate the primal variables in terms of the Lagrange multipliers. Newton Raphson is thus used to solve the system
![Page 11: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/11.jpg)
11
Newton-Raphson iteration (Taig & Kerr, 1973)
Iteration scheme on Lagrange multipliers only
Gradient matrix H is given by
![Page 12: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/12.jpg)
12
Newton-Raphson iteration (Taig & Kerr, 1973)
Difficulties
– Select an appropriate initial dual set (0)
– Find correct set of active / passive design variables
– Identify the set of estimated active behavior constraints (i.e. nonzero j’s)
– H might become singular at some stage of the process
Solution: dual methods ➔ Generalized Optimality Criteria
– H is indeed the Hessian matrix of the dual function
Dual maximization: determination of the Lagrange variable values
Set of active constraints = Non zero optimum values of the Lagrange multipliers
![Page 13: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/13.jpg)
13
Ten-bar-truss example
The stress-ratioïng itself tends to increase the design variable with the smallest stress limit
Example: stress limit = 25000psi except in member 8 with a variable limit from 25000 to 70000 psi
![Page 14: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/14.jpg)
14
Stress and displacement constraints
![Page 15: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/15.jpg)
GENERALIZED OPTIMALITY CRITERIA➔
DUAL SOLUTION OF SUBPROBLEMS
15
![Page 16: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/16.jpg)
Optimality Criteria
Explicit optimization problem:
Expression of the displacement constraints using the virtual force method
Cij are constant in statically determinate structures
16
![Page 17: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/17.jpg)
Optimality Criteria
Explicit optimality conditions of the minimum = KKT conditions
That is
➔ Analytical expression of the design variables in terms of the
Lagrange variables (primal dual relationships)
17
![Page 18: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/18.jpg)
Optimality Criteria
Non negativity constraints of the Lagrange multipliers
– Active constraints
– Passive constraints
Optimal j* ➔ optimal xi*
Lagrange multipliers = new variables
= dual variables
18
![Page 19: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/19.jpg)
Generalized Optimality Criteria
DUAL MAXIMIZATION
Identifying the optimal value of the Lagrange multipliers j by solving the dual problem
With the dual function
Primal dual relationships
19
![Page 20: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/20.jpg)
Dual methods
Primal problem
Lagrange function of the problem
KKT conditions
20
![Page 21: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/21.jpg)
Dual methods
Primal dual relations
Because of the separabilty: solution of n 1D problems
Gives the explicit relation of primal dual variables
21
![Page 22: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/22.jpg)
Dual methods
Dual function
Dual problem
22
![Page 23: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/23.jpg)
Dual methods
PROPERTIES OF DUAL FUNCTION: Derivative of dual function
Optimality conditions of the dual function
23
![Page 24: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/24.jpg)
Dual methods
Algorithms for maximizing the dual function based on the Newton’s methods = Rigorous Update procedure of the Lagrange Multipliers
Ascent direction
Hessian of the dual function
➔ Discontinuous because of the modification of ñ, the set of active
design variables 24
![Page 25: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/25.jpg)
Dual methods
Second order discontinuous planes:
25
![Page 26: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/26.jpg)
Generalized Optimality Criteria
Dual problems:
– Explicit
– Quasi unconstrained
– Gradient directly available
Dual algorithms
– Low computational cost (like OC techniques)
– Reliable (mathematical basis)
– Can handle large numbers of inequality constraints
– Automatically find the active set of constraints and the corresponding active passive design variables
Easy to solve by standard methods
26
![Page 27: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/27.jpg)
Dual methods
Two-bar truss
27
![Page 28: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/28.jpg)
Dual methods
GOC relations
KKT conditions ➔ primal dual relations
Discontinuity planes
28
![Page 29: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/29.jpg)
Dual methods
Explicit dual function for 2-bar truss
29
![Page 30: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/30.jpg)
Dual methods
30
![Page 31: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/31.jpg)
Dual methods
31
![Page 32: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/32.jpg)
Dual methods
Line search
32
![Page 33: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/33.jpg)
Dual methods
10-bar truss – Conventional andGeneralized OC
33
![Page 34: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/34.jpg)
CONCLUSION
34
![Page 35: FROM OPTIMALITY CRITERIA TO SEQUENTIAL CONVEX PROGRAMMING](https://reader031.fdocuments.us/reader031/viewer/2022021804/620d919ab4c4c50d5748da1a/html5/thumbnails/35.jpg)
CONCLUSION
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
Sequence of explicit subproblems
– First Order Approximations
Efficient solution of optimization problem
Primal / dual solution schemes
35