1 / 33

Abhishek K. Shrivastava September 25 th , 2009

Listing Unique Fractional Factorial Designs – I. Abhishek K. Shrivastava September 25 th , 2009. Outline. Fractional Factorial Designs (FFD). What are experiments & designs? What are FFDs? Why is there a list? Are there many FFDs?.

Download Presentation

Abhishek K. Shrivastava September 25 th , 2009

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Listing Unique Fractional Factorial Designs – I Abhishek K. Shrivastava September 25th, 2009

  2. Outline • Fractional Factorial Designs (FFD) What are experiments & designs?What are FFDs? Why is there a list? Are there many FFDs? Design isomorphismListing designs Listing unique designs – brute force gen • Listing Unique designs • Graphs & designs What are graphs?FFDs as graphs • FFDI & GI Solving GI – canonical labeling (nauty)Implications to generating design catalogs Abhishek K. Shrivastava, TAMU

  3. Experiments, Designs & Fractional factorial designs (FFDs)

  4. Experiments • Experiments for quantifying effect of causal variables Effect of process parameters on product quality Source: http://www.emeraldinsight.com/fig/0680170207035.png Miller-Urey Experiment Source: http://www.physorg.com • Experiments for testing hypothesis Abhishek K. Shrivastava, TAMU

  5. Analyze datay = Xb+e Experimental Designs • Choose variable settings to collect data • Replicate runs • Randomize run order Collect Data Experimental plan Experimental design Make inferences Abhishek K. Shrivastava, TAMU

  6. Experimental Designs factors a run Levels of factor I Experimental design Abhishek K. Shrivastava, TAMU

  7. Experiments with 5 factors • Suppose each factor has 2 runs Choice of design? • Full factorial, i.e. 25 = 32 runs • Too many runs (2n) • Fractional factorial design (FFD) • Pick some subset of full factorial runs • Many fractional factorial designs exist • 25–2 design with 8 runs • Generated using defining relations D=BC and E=AB (regular FFD) Abhishek K. Shrivastava, TAMU

  8. Listing FFDs • Using FFDs • Reduces experimenter’s effort • But at a cost! • Hypothetical example: 25–2 design with D=A, E=AB • Can estimate effect of A+D • Many different FFDs with different statistical capability • How do you choose an FFD?? Abhishek K. Shrivastava, TAMU

  9. Design catalogs • Catalog of 16-run regular FFDs (Wu & Hamada, 2000) • Compare statistical properties to choose Issues: • Large size regular FFDs not available? • Other classes of FFDs not available Abhishek K. Shrivastava, TAMU

  10. Listing Unique FFDs

  11. Unique designs: 7-factor FFD example • 7 factors: • Cutting speed . . . . . . . . • Feed . . . . . . . . . . . . • Depth of cut . . . . . . . . • Hot/cold worked work piece . • Dry/wet environment . . . . • Cutting tool material . . . . . • Cutting geometry . . . . . . A B C D E F G Abhishek K. Shrivastava, TAMU

  12. Unique designs: 7-factor FFD example • 7 factors: • Cutting speed . . . . . . . . • Feed . . . . . . . . . . . . • Depth of cut . . . . . . . . • Hot/cold worked work piece . • Dry/wet environment . . . . • Cutting tool material . . . . . • Cutting geometry . . . . . . . A B C D E F G A C B D F E G (a) Defining words: {ABE, ACF, BDG} (b) Defining words: {ABE, ACF, CDG} Abhishek K. Shrivastava, TAMU

  13. Unique designs Reordered matrix, exchanged columns B↔C, E↔F, reordered rows in (a) (b) Defining words: {ABE, ACF, CDG} Abhishek K. Shrivastava, TAMU

  14. Unique designs • Designs (a) & (b) • are isomorphic under factor relabeling & row reordering • have same statistical properties (a) Defining words: {ABE, ACF, BDG} (b) Defining words: {ABE, ACF, CDG} Abhishek K. Shrivastava, TAMU

  15. FFD Isomorphism (FFDI) • Definition. Two FFD matrices are isomorphicto each other if one can be obtained from the other by • some relabeling of the factor labels, level labels of factors and row labels. • FFDI problem. Computational problem of determining if two FFDs are isomorphic. Abhishek K. Shrivastava, TAMU

  16. Design catalogs • No two designs should be isomorphic • Non-isomorphic catalogs • Why? • Isomorphic designs are statistically identical • Discarding isomorphs can drastically reduce catalog size • e.g., # 215–10 designs > 5 million, where # unique (i.e, non-isomorphic) designs is only 144! Abhishek K. Shrivastava, TAMU

  17. add column/ factor add column/ factor add column/ factor add column/ factor … 24 Full factorial 5-factor FFD 6-factor FFD 7-factor FFD Listing Unique FFDs • Consider 16-run designs – sequential generation • How do you pick these columns?? FFD class • Regular FFD: defining relation E=AB, F=AC, G=BD • Orthogonal arrays: added column keeps orthogonal array property • All possible choices of columns gives the catalog Abhishek K. Shrivastava, TAMU

  18. discard isomorphs 7-factor designs from 6-factor designs ... 24 design Non-isomorphic 5-factor designs Non-isomorphic 6-factor designs Non-isomorphic 7-factor designs Intermediate step Listing Unique FFDs • Consider sequential generation of 16-run designs • Note: reducing # intermediate designs will speed up the algorithm • How to discard isomorphs? Abhishek K. Shrivastava, TAMU

  19. Solving FFDI: literature review Two types of tests in literature • Necessary checks • faster • Word length pattern, letter pattern matrix, centered L2 discrepancy, extended word length pattern, moment projection pattern, coset pattern matrix • Necessary & Sufficient checks • slower / computationally expensive • exhaustive relabeling, Hamming distance based, minimal column base, indicator function representation based, eigenvalues of word pattern matrices (conjectured) • Legend: • Regular FFDs only • All FFDs Fastest; 2-level regular FFDs only Abhishek K. Shrivastava, TAMU

  20. … Proposed FFDI solution (in a nutshell) • Graph models for FFDs • Equivalence between FFDI and GI • Solving GI Construct graphs from FFDs Solve graph isomorphism problem FFD class specific Abhishek K. Shrivastava, TAMU

  21. Graphs and FFDs • Graphs & Graph isomorphism • 2-level regular FFDs • Multi-level regular FFDs • Non-regular FFDs • 2-level regular split-plot FFDs

  22. Some 2-level regular FFD terminology • Defining relations: E=AB, F=AC, G=BD • E=AB E=(A+B) mod 2 • (A+B+E) mod 2 = ABE = I (identity) • Defining words: ABE, ACF, BDG • Other words (by mod-2 sum), e.g., BCEF (= ABE+ACF) • Defining contrast subgroup – all words generated from defining words • S = {I, ABE, ACF, BDG, BCEF, ADEG, ABCDFG, CDEFG} A regular 27–3 design Abhishek K. Shrivastava, TAMU

  23. 2-level regular FFD isomorphism (rFFDI) • Two regular FFDs, represented by their defining contrast subgroups S1, S2 are isomorphicto each other iff • one of S1 or S2 can be obtained from the other by some permutation of factor labels and reordering of words. • Example: two 7-factor designs, S1 = {I, ABE, ACF, BDG, BCEF, ADEG, ABCDFG, CDEFG}, S2 = {I, ABE, ACF, CDG, BCEF, ADFG, ABCDEG, BDEFG} S1 = {I, ABE, ACF, BDG, BCEF, ADEG, ABCDFG, CDEFG} B↔CE↔F S1' = {I, ACF, ABE, CDG, CBFE, ADFG, ACBDEG, BDFEG} rewrite S1' = {I, ABE, ACF, CDG, BCEF, ADFG, ABCDEG, BDEFG} S2 Abhishek K. Shrivastava, TAMU

  24. 2-level regular FFDs as bipartite graphs Example: n = 7, S = {I, ABE, ACF, BDG, BCEF, ADEG, ABCDFG, CDEFG} • Start with G(V,E) = empty graph (no vertices); V = VaVb • For each factor in d, add a vertex in Va • For each word in S, except I , add a vertex in Vb • For each word in S, except I , add edges between the word’s vertex (in Vb) and the factors’ vertices (in Va) Abhishek K. Shrivastava, TAMU

  25. 2-level regular FFD isomorphism problem Bipartite graph isomorphism Bipartite graph isomorphism • [Bipartite graph isomorphism problem] Given two graphs, does there exist a graph isomorphism that preserves vertex partitions. • Is GI-complete • Same computational complexity as GI • FFD to Graph conversion takes O(n|S|)steps Abhishek K. Shrivastava, TAMU

  26. Multi-level designs as Multi-graphs • Multi-graph representation of a 35–2 design with defining contrast subgroup {I, ABCD2, A2B2C2D, AB2E2, A2BE, AC2DE, A2CD2E2, BC2DE2, B2CD2E} • Similar representation for mixed level designs Abhishek K. Shrivastava, TAMU

  27. Non-regular designs as Vertex-colored graphs • Vertex-colored graph representation A 4-factor, 5-run design *edges colored only for better visualization Abhishek K. Shrivastava, TAMU

  28. Non-regular FFD isomorphism problem Vertex colored graph isomorphism Vertex-colored graph isomorphism • [Vertex colored graph isomorphism problem] Given two graphs, does there exist a graph isomorphism that preserves vertex colors. • Is GI-complete • Same computational complexity as GI Abhishek K. Shrivastava, TAMU

  29. 2-level regular split-plot FFD (FFSP) • FFDs with restricted randomization of runs • Turning part quality example • Cutting speed (A), depth of cut (B), feed (C) is not to be changed after every run • Two groups of factors • Whole plot factors: difficult to change, e.g., A, B, C in above example • Sub-plot factors: easy to change, e.g., d, e, f and g in above example • Relabeling A ↔ d not permitted anymore Abhishek K. Shrivastava, TAMU

  30. Regular FFSPs • Regular fractional factorial designs with restricted randomization • Uniquely represented by defining contrast subgroup • e.g., 2(3–1)+(4–2) design with C=AB, f=de, g=Bd • Defining relations for whole plot factors have no sub-plot factors, e.g., C=AB • Defining relations for sub-plot factors have at least one sub-plot factor A 2(3–1)+(4–2) design matrix Abhishek K. Shrivastava, TAMU

  31. FFSP Isomorphism • [Definition V.1] Two FFSP matrices are isomorphicto each other if one can be obtained from the other by • some relabeling of the whole-plot factor labels, sub-plot factor labels, level labels of factors and row labels. • [Proposition V.2] Two FFSPs, represented by their defining contrast subgroups S1, S2 are isomorphicto each other iff • one of S1 or S2 can be obtained from the other by some permutation of whole-plot factor labels and sub-plot factor labels, and reordering of words. Abhishek K. Shrivastava, TAMU

  32. FFSPs as vertex-colored graphs • Vertex-colored graphs • Each vertex has color • Graph construction • Similar to regular FFDs • Whole-plot factors, sub-plot factors, words – all have different colors • Other variants: split-split-plot designs, non-regular split-plot designs Abhishek K. Shrivastava, TAMU

  33. GI and FFDI • Solving GI: canonical labeling • Implications to listing FFDs efficiently …next week

More Related