Mechanical Engineering Department Non-Probabilistic Design Optimization with Insufficient Data using Possibility and Evidence Theories Zissimos P. Mourelatos Jun Zhou Mechanical Engineering Department Oakland University Rochester, MI 48309, USA mourelat@oakland.edu
Overview Introduction Design under uncertainty Uncertainty theories Possibility – Based Design Optimization (PBDO) Uncertainty quantification and propagation Design algorithms Evidence – Based Design Optimization (EBDO) Examples Summary and conclusions
Design Under Uncertainty Analysis / Simulation Input Output Uncertainty (Quantified) Uncertainty (Calculated) Propagation Design Quantification 2. Propagation 3. Design
Uncertainty Types Aleatory Uncertainty (Irreducible, Stochastic) Probabilistic distributions Bayesian updating Epistemic Uncertainty (Reducible, Subjective, Ignorance, Lack of Information) Fuzzy Sets; Possibility methods (non-conflicting information) Evidence theory (conflicting information)
Uncertainty Theories Evidence Theory Probability Theory Possibility Theory Probability Theory
Non-Probabilistic Design Optimization: Set Notation Universe (X) Power Set (All sets) Element A B C Evidence Theory
Possibility-Based Design Optimization (PBDO)
Possibility-Based Design Optimization (PBDO) Evidence Theory No Conflicting Evidence (Possibility Theory)
Quantification of a Fuzzy Variable: Membership Function - cut provides confidence level At each confidence level, or -cut, a set is defined as convex normal set
Propagation of Epistemic Uncertainty Extension Principle The “extension principle” calculates the membership function (possibility distribution) of the fuzzy response from the membership functions of the fuzzy input variables. If where then Practical Approximations of Extension Principle Vertex Method Discretization Method Hybrid (Global-Local) Optimization Method
Optimization Method Global where : s.t. and s.t. 1.0 1.0 a a 0.0 0.0
Possibility-Based Design Optimization (PBDO) (Possibility Theory) What is possible may not be probable What is impossible is also improbable If feasibility is expressed with positive null form then, constraint g is ALWAYS satisfied if for or
Possibility-Based Design Optimization (PBDO) Considering that , we have
Possibility-Based Design Optimization (PBDO) s.t. ; ; s.t. OR Double Loop
PBDO with both Random and Possibilistic Variables , with Triple Loop ; , s.t.
Evidence-Based Design Optimization (EBDO)
Evidence-Based Design Optimization (EBDO) Basic Probability Assignment (BPA): m(A) If m(A)>0 for then A is a focal element 0.2 0.1 0.4 0.3 Y 5 6.2 8 9.5 11 “Expert” A 5 0.3 0.4 Y 7 8.7 11 “Expert” B Y 0.x1 7 8.7 11 9.5 8 6.2 0.x2 0.x3 0.x6 0.x5 0.x4 5 Combining Rule (Dempster – Shafer)
Evidence-Based Design Optimization (EBDO) Assuming independence, For where define: Assuming independence,
Evidence-Based Design Optimization (EBDO) BPA structure for a two-input problem
Evidence-Based Design Optimization (EBDO) Uncertainty Propagation If we define, then where and
Evidence-Based Design Optimization (EBDO) Position of a focal element w.r.t. limit state Contributes to Belief Contributes to Plausibility
Evidence-Based Design Optimization (EBDO) If non-negative null form Design Principle If non-negative null form is used for feasibility, feasible infeasible failure Therefore, is satisfied if OR
Evidence-Based Design Optimization (EBDO) Formulation ,
Evidence-Based Design Optimization (EBDO) Calculation of
Geometric Interpretation of PBDO and EBDO
Possibility-Based Design Optimization (PBDO) Implementation x1 initial design point Feasible Region g1(x1,x2)=0 frame of discernment g2(x1,x2)=0 Objective Reduces PBDO optimum deterministic optimum x2
Evidence-Based Design Optimization (EBDO) deterministic optimum Implementation Feasible Region x2 x1 g1(x1,x2)=0 g2(x1,x2)=0 Objective Reduces initial design point frame of discernment EBDO optimum deterministic optimum
Evidence-Based Design Optimization (EBDO) deterministic optimum Implementation Feasible Region x2 x1 g1(x1,x2)=0 g2(x1,x2)=0 Objective Reduces hyper-ellipse initial design point frame of discernment B MPP for g1=0 deterministic optimum
Evidence-Based Design Optimization (EBDO) deterministic optimum Implementation Feasible Region x2 x1 g1(x1,x2)=0 g2(x1,x2)=0 Objective Reduces hyper-ellipse initial design point frame of discernment EBDO optimum B MPP for g1=0 deterministic optimum
Cantilever Beam Example: RBDO Formulation s.t. where :
Cantilever Beam Example: PBDO Formulation s.t. RBDO s.t. PBDO
Cantilever Beam Example: EBDO Formulation s.t. EBDO BPA Structure
Cantilever Beam Example: EBDO Formulation BPA structure for y, Y, Z, E
Cantilever Beam Example: Comparison of Results
Thin-walled Pressure Vessel Example yielding s.t.
Thin-walled Pressure Vessel Example BPA structure for R, L, t, P and Y
Thin-walled Pressure Vessel Example
More Conservative Design Summary and Conclusions Possibility and evidence theories were used to quantify and propagate uncertainty. PBDO and EBDO algorithms were presented for design with incomplete information. EBDO design is more conservative than the RBDO design but less conservative than PBDO design. Deterministic RBDO EBDO PBDO Less Information More Conservative Design
Q & A