1. 1 Andrea Saltelli, Jessica Cariboni and Francesca Campolongo European Commission, Joint Research Centre SAMO 2007 Budapest Accelerating factors screening

2. 2 1.Sensitivity analysis web at JRC (software, tutorials,..) http://sensitivity-analysis.jrc.cec.eu.int/ 2.New book on SA with exercises for students - at Wiley for review - Please flag errors! 3.Summer school in 2008 – date to be decided Sensitivity analysis at the Joint Research Centre of Ispra

3. 3 Where do we stands in terms of good practices for global SA : Screening: Morris – Campolongo – EE (1991-2007) Quantitative: Sobol’, plus several investigators, 1990- 2007

4. 4 Screening: Morris – Campolongo – EE (1991- 2007) Good but not so efficient Quantitative: Sobol’, Saltelli (1993-2002) Efficient for S i (Mara’ + Tarantola [scrambled FAST], Ratto + Young [SDR] + proximities [Marco’s presentation of yesterday]) Not so efficient for S Ti (Saltelli 2002)

5. 5 The EE method can be seen as an extension of a derivative-based analysis. Where to start? From the best available practice in screening: The method of Elementary Effects (Morris 1991) Max Morris, Department of Statistics Iowa State University

6. 6 The method of Elementary Effects Model Elementary Effect for the i th input factor in a point X o

7. 7 r elem. effects EE 1 i EE 2 i … EE r i are computed at X 1, …, X r and then averaged. Average of EE i ’s   (x i ) Standard deviation of the EEi’s   (xi) Factors can be screened on the  (xi)  (xi) plane Using EE method: The EEi is still a local measure Solution: take the average of several EE

8. 8 A graphical representation of results

9. 9 Using the EE method Each input varies across p possible values (levels – quantiles usually) within its range of variation xi U(0,1) p = 4  p 1 = 0 p 2 = 1/3 p 3 = 2/3 p 4 = 1 The optimal choice for  is  = p / 2 (p -1) 01/32/3101/ 3 2/31 Grid in 2D Sampling the levels uniformly

10. 10 Improving the EE (Campolongo et al., ….. 2007) - Taking the modulus of  (xi),  *(xi) Instead of using the couple of  (xi) and  (xi) x1x1 x2x2 AB C A’ C’B’ -Maximizing the spread of the trajectories in the input space -Application to groups of factors

11. 11 S Ti available analytically a=99 a=9 a=0.9 A comparison with variance-based methods: Is  *(xi) related to either S i or S Ti ? Empirical evidence: the g-function of Sobol’

12. 12 Empirical evidence: the g-function Factor a(i) x10.001 x289.9 x35.54 x442.1 x50.78 x61.26 x70.04 x80.79 x974.51 x104.32 x1182.51 x1241.62 A comparison with variance-based   *(x i ) is a good proxy for S Ti

13. 13 Implementing the EE method Original implementation estimate r EE’s per input. r trajectories of (k+1) sample points are generated, each providing one EE per input A trajectory of the EE design Total cost = r (k + 1) r is in the range 4 -10 Each trajectory gives k effect EE at the cost of (k + 1) simulations. Efficiency =k/(k+1)~1

14. 14 Conclusion: the EE is a useful method Is its efficiency k/(k+1) ~ 1 good? We can compare with the Saltelli 2002 method to implement the calculation of the first order and total order sensitivity indices:

15. 15 One of this plus … … one of this plus … plus K of these With: One can compute all first and total effects for k factors Saltelli 2002

16. 16 One of this K of these Total: N(K+2) runs To obtain N*2*k elementary effects (for S i or S Ti ) Efficiency=2k/(k+2)~2 Better that the EE method. Saltelli 2002

17. 17 Conclusion: the efficiency of EE might have scope for improvement. The better efficiency of the global method (Saltelli 2002) against the screening method (EE) is due to the fact that two effects (one of the first order and one of the total order) are computed from each row of Ai. Can we do the same with EE?

18. 18 … is one step in the non- X i direction (all moves but X i ) Saltelli 2002 From To

19. 19 … is one step in the X i direction (X i moves and X ~i does not) Saltelli 2002 From To

20. 20 How about alternating steps along the X i ’s axes with steps along the along the X ~i ’s also for an EE-line screening method? How can we combine steps along X i ’s axes with steps along the X ~i ’s?

21. 21 Can we generate efficiently exploration trajectories in the hyperspace of the input factors where steps in the X i and X ~i directions are nicely arranged, e.g. in a square? Beyond Elementary Effects Method

22. 22 Beyond Elementary Effects Method

23. 23 Our thesis is that (1) Both |y 1 -y 3 | and |y 2 -y 4 | tells me about the first order effect of X 1

24. 24 … and that : (2) ||y 1 -y 4 |-|y 1 -y 2 ||, ||y 2 -y 3 |-|y 2 -y 1 ||, ||y 3 -y 2 |-|y 3 -y 4 ||, ||y 4 -y 1 |-|y 4 -y 3 ||, all tell me about the total order effect of a factor.

25. 25 Before trying to substantiate our thesis we give a look at how these squares could be built efficiently Four runs, six factors

26. 26 Four runs, six factors, six steps along the X ~i directions We call these four runs ‘base runs’

27. 27 Base runs Clones For each step in the X ~i direction we add two in the X i direction

28. 28 Base runs Clones Let’s count: Run 3 is a step away from run 1 in the X 1 direction. Run 4 is a step away from run 2 in the X 1 direction. Run 2 was already a step away from run 1 in the X ~1 direction Run 4 is also a step away from run 3 in the X ~1 direction … the square is closed.

29. 29 Beyond Elementary Effects Method

30. 30 Base runs Clones Let do some more counting. We have 4 base runs, 16 runs in total, six factors and four effects for factor. Efficiency= 24/16=3/2

31. 31 For 6 base runs, we have 15 factors, 36 runs in total, again four effects for factor. Efficiency= 60/36 ~ 2 for increasing number of factors … It would be nice to stop here! … but let us go back to the 6 factors example

32. 32 There are many more effects hidden in the scheme: e.g. three more effects for run 16. Most of these effects are of the X ~i type The number of extra terms is between 2k and 4 k

33. 33 The number of extra terms grows with k Some of these need only one more point to close a square Most of these need two extra points to close a square

34. 34 Let us forget about the additional terms for the moment and let us try screening …

35. 35 Numerical Experiment: g-function where

36. 36 Results: g-function (180 runs) a1a1 a2a2 a3a3 a4a4 a5a5 a6a6 a7a7 a8a8 a9a9 a 10 a 11 a 12 a 13 a 14 a 15 0.010.020.05990.301.50785789960.5098878890

37. 37 Number of runs: EE(2007)= 25; EE =22 K=10, a=(0.01,0.02,0.015,99,78,57,89,97,96,87)

38. 38 Test function Book (2007) The last two Z’s and the last two omegas are the most important factors

39. 39 Number of runs: new method= 64; old method =58

40. 40 g function 25 replicas of EE1(2007)

41. 41 g function 25 replicas of EE2(2007)

42. 42 g function 25 replicas of EE

43. 43 book function 25 replicas of EE2(2007)

44. 44 book function 25 replicas of EE1(2007)

45. 45 book function 25 replicas of EE

46. 46 What next? Good for S i, S Ti ?

47. 47 S i couple S Ti couple S i couple S Ti couple

48. 48 S i couple S Ti couple S i couple S Ti couple Try to exploit this design for the improvement of the Saltelli 2002 method for the S Ti

49. 49 The number of extra terms grows with k Some of these need only one more point to close a square Most of these need two extra points to close a square (closed squares give 4 effects, 2 S i & 2 S Ti )

50. 50 Conclusions The new scheme (aka il matricione ) has promises for EE and S Ti Work on the algorithms is needed to make a sizeable difference with best available practices …

51. 51 il matricione