Helsinki University of Technology Systems Analysis Laboratory RICHER
Helsinki University of Technology Systems Analysis Laboratory RICHER – A Method for Exploiting Incomplete Ordinal Information in Value Trees Antti Punkka and Ahti Salo Systems Analysis Laboratory Helsinki University of Technology P. O. Box 1100, 02015 HUT, Finland forename. surname@hut. fi http: //www. sal. hut. fi/
Helsinki University of Technology Systems Analysis Laboratory Value tree analysis n n m alternatives: X={x 1, …, xm} , n attributes: A={a 1, …, an} Additive value function non-normalized form or normalized form 2
Helsinki University of Technology Systems Analysis Laboratory Preference elicitation n Complete information – Point estimates, e. g. w 1=0. 5 – E. g. , SMART (Edwards 1977) n Incomplete information: preference programming methods – Weight ratio and weight intervals » – Intervals for normalized scores » – PAIRS (Salo and Hämäläinen 1992), PRIME (Salo and Hämäläinen 2001), Arbel’s approach (1989) n Ordinal information – Rank attributes in terms of relative importance » point estimates through, e. g. , rank sum weights (Stillwell et al. 1981) » incomplete ordinal information (RICH; Salo and Punkka 2004) 3
Helsinki University of Technology Systems Analysis Laboratory Incomplete preference information n Complete information hard to acquire – Relative importance of attributes – Alternatives’ properties – Incomplete information » n Overall value intervals for alternatives – Smallest and largest possible value from LP where Sw is the feasible region for the attribute weights 4
Helsinki University of Technology Systems Analysis Laboratory Pairwise dominance relation n Alternative xk dominates xj in the sense of pairwise dominance – Two attributes, and positive with some feasible scores and weights V n Several alternatives may remain nondominated – Additional preference statements to make the feasible region smaller – Decision rules assist the DM in selection of the most preferred one x 1 dominates x 2 w 1 w 2 0. 4 0. 6 0. 7 0. 3 5
Helsinki University of Technology Systems Analysis Laboratory Incomplete ordinal preference information n Complete ordinal information is a complete rank-ordering of attributes or alternatives – Rankings are exactly known for each alternative – Leads to a convex set of feasible scores and weights, when interpreted as incomplete preference information n The RICH (Rank Inclusion in Criteria Hierarchies) method – – Incomplete ordinal statements about relative importance of attributes ”Cost is the most important attribute” ”Environmental factors is among the three most important attributes” Several rank-orderings can be compatible with the preference statements » e. g. : either attribute a 1 or a 2 is the most important of the three attributes n a 3 is either the second or the least important one » may lead to non-convex feasible region of the attribute weights 6
Helsinki University of Technology Systems Analysis Laboratory Non-convex feasible region in RICH n n ”Either a 1 or a 2 is the most important of the three attributes” Calculation by dividing into compatible rank-orderings – Extreme points readily computed – Lower bounds for weights wi b 0 n Full support provided by RICH Decisions ©, http: //www. decisionarium. hut. fi/ – Applications » evalution of risk management tools (Ojanen et al. 2004) » support for setting priorities for a research programme in wood material science (Salo and Liesiö 2004) 7
Helsinki University of Technology Systems Analysis Laboratory The RICHER (RICH with Extended Rankings) method n Extends incomplete ordinal information to alternatives – ”Alternatives x 1, x 2 and x 3 are three most preferred with regard to environmental factors” – ”Alternative x 1 is not among the three most preferred ones” – ”Considering alternatives x 1, x 2 and x 3, the least preferred with regard to cost is x 1” n Statements about attribute weights incorporated as well n Comparison to the RICH method – – – Suitable also for statements about alternatives Computationally much more efficient Includes all features of RICH Allows evaluation within subsets Applicable in conjunction with other preference programming methods 8
Helsinki University of Technology Systems Analysis Laboratory Modeling of incomplete ordinal information (1/4) n Rank-ordering –function r – Bijection from (sub)set of alternatives X’ X (or (sub)set of attributes) to set of rankings – E. g. , r=(r(x 1), r(x 2), r(x 3))=(1, 3, 2) – The smaller the ranking, the better the alternative » e. g. , ”r(x 4)=1 the ranking of x 4 is 1, i. e. it is the most preferred” – Several rank-orderings may be compatible with the preference information – Incomplete ordinal statements about alternatives can be expressed with regard to different sets of attributes A’ » single attribute, (sub)set of attributes or holistic statements considering all attributes » e. g. , one can subject statements to cost and environmental factors together 9
Helsinki University of Technology Systems Analysis Laboratory Modeling of incomplete ordinal information (2/4) n Non-normalized form of value function n Set of feasible values V includes score vectors v=(v(x 1), . . . , v(xm)) – v(xk) denotes the value of xk with regard to some set of attributes » e. g. , if A’={a 2}, then v(xk)=v 2(x 2 k) » e. g. , if A’=A, then v(xk)=V(xk) – Restricted by preference statements n Feasible region associated with a rank-ordering is convex 10
Helsinki University of Technology Systems Analysis Laboratory Modeling of incomplete ordinal information (3/4) n Elicitation of the preference statements is carried out through an alternative set I X’ X and a ranking set J {1, . . . , m’}, where m’=|X’| – If |I| |J|, the rankings in J are attained by alternatives in I – If |I|<|J|, the alternatives in I have their rankings in J – Sets subjected to X’ and A’ denoted by I(A’, X’) and J(A’, X’) n Examples – x 1 and x 2 are among the three most preferred ones with regard to cost attribute a 1. Now A’={a 1} and X’=X, I({a 1}; X)={x 1, x 2}, J({a 1}; X)={1, 2, 3} – Holistically (A’=A) the two least preferred are among x 4, x 5, x 8, x 9: I(A; X)= {x 4, x 5, x 8, x 9}, J(A; X)={m-1, m} – Holistically the most preferred of the set X’={x 1, x 2, x 7} is x 1: I(A; X’)={x 1}, J(A; X’)={1} 11
Helsinki University of Technology Systems Analysis Laboratory Modeling of incomplete ordinal information (4/4) n Sets I and J lead to compatible rank-orderings R(I, J) n The feasible region associated with many compatible rankorderings is usually non-convex n Statements can be given with regard to different attribute sets – Several rank-orderings may be compatible with each of these sets 12
Helsinki University of Technology Systems Analysis Laboratory Mixed integer linear programming model (1/5) n Overcoming the non-convexity – Continuous ”milestone variable” zk distinguishes between the values of alternatives with rankings k and k+1 – If xj’s ranking is at most k, its value is at least zk and we let yk(xj)=1, else 0 – There are exactly k alternatives whose ranking is at most k » e. g. , the three rankings 1, 2 and 3 are at most 3 13
Helsinki University of Technology Systems Analysis Laboratory Mixed integer linear programming model (2/5) n Formally n For the sake of interpretational and computational matters, we set 14
Helsinki University of Technology Systems Analysis Laboratory Mixed integer linear programming model (3/5) n Adding preference statements into the model n Assumption |J|≤|I| – For all rankings j J, the respective alternative belongs to I n Because of the uniqueness of the rankings, there is exactly one alternative in I, for which yj-1(xi)=0, and yj(xi)=1. For other alternatives, yj-1(xi) and yj(xi) get same values n E. g. , I=(x 1, x 2, x 5), J={2, 4}, exactly one of the alternatives has the ranking 2 it is the only one with different values for y 2(xi) and y 1(xi) 15
Helsinki University of Technology Systems Analysis Laboratory Mixed integer linear programming model (4/5) n Some milestone and binary variables and the respective constraints are redundant – Given a statement that alternatives x 1 and x 2 are the two most preferred, for example variables z 1, z 3 and y 1(xj), y 3(xj) are not needed » actually only z 2 and y 2(xj) are needed n If set J is ”sequential”, i. e. , it constitutes of consecutive positive integers, the number of variables and constraints can be substantially decreased – For example sets {3, 4, 5} and {1} are sequential, set {1, 3} not n For the compatible rank-orderings associated with sets I and J, |I| |J|, it holds 16
Helsinki University of Technology Systems Analysis Laboratory Mixed integer linear programming model (5/5) n Partitioning of J into a minimal number of sequential sets Ji – For example, J={1, 2, 6, 7} is partitioned into J 1={1, 2} and J 2={6, 7} – At most 2 milestone and 2 m’ binary variables needed to represent the statement associated to a sequential set Ji n Representation of the feasible region S(I, J) as the intersection of the feasible regions S(I, Ji) – Constraints for all pairs I, Ji are set in the same model n If contradictionary to the assumption it holds |J|>|I|, the feasible region is constructed with the help of complement sets IC=X’I and JC={1, . . . , m’}J – Now |JC|≤|IC| – S(I, J)=S(IC, JC) n All linear inequalities can be included in the model 17
Helsinki University of Technology Systems Analysis Laboratory An illustrative example (1/6) n A company is about to choose the facility for a new office – Eight attributes relevant: a 1: size of the office, a 2: rental costs, a 3: renovation need, a 4: car park opportunities, a 5: means of communication, a 6: distance to city center, a 7: other facilities in the neighborhood, a 8: habitability – 12 alternatives 18
Helsinki University of Technology Systems Analysis Laboratory An illustrative example (2/6) n Attributes size, rental costs, car park opportunities and distance to city center are assessed through [0, 1]-normalized scores or score intervals 19
Helsinki University of Technology Systems Analysis Laboratory An illustrative example (3/6) n Other information is turned into incomplete ordinal statements – E. g. , alternative x 2 is the only one with no renovation need (a 3), hence the ranking 1 – E. g. , alternative x 2 is the least preferred w. r. t. habitability (a 8) 20
Helsinki University of Technology Systems Analysis Laboratory An illustrative example (4/6) n Information on attributes’ relative importance – Complete rank-ordering of the attributes is r(a 1, a 2, . . . , a 8)=(1, 2, . . . , 8) – A weight of 0. 50 is assigned to the most important attribute, size of the office – Weights are lower bounded by wi 1/3 n n A holistic preference for x 4 over x 1 over x 3 21
Helsinki University of Technology Systems Analysis Laboratory An illustrative example (5/6) n Pairwise bounds (minima of overall value differences) indicate that there are 5 non-dominated alternatives x 5, x 7, x 8, x 9 and x 10 22
Helsinki University of Technology Systems Analysis Laboratory An illustrative example (6/6) n To discriminate between non-dominated alternatives, decision rules are applied n Each rule recommends alternative x 5 n XPress-MP was used in solving the example n Calculation of the example (pairwise bounds, overall value intervals for each 12 alternatives; 156 MILPs) took 14 seconds on a Pentium III at 800 MHz with 256 MB RAM 23
Helsinki University of Technology Systems Analysis Laboratory Conclusion n The DM can give incomplete ordinal information about the alternatives with regard to a single attribute, a set of attributes or holistically n Statements about relative importance of attributes are allowed, as well n Based on a linear model and hence it can be used in conjunction with other preference programming methods n Computationally far more efficient than RICH, and more flexible as it contains all features of RICH n Software implementation of RICHER Decisions © ongoing n Future research directions – Modeling of classification procedures with RICHER methodology (cf. the example in this presentation) – Application of RICHER methodology to voting or other group decision processes – Application of incompelete ordinal information in Robust Portfolio Modeling (RPM) 24
Helsinki University of Technology Systems Analysis Laboratory Related references Arbel, A. , “Approximate Articulation of Preference and Priority Derivation”, European Journal of Operations Research 43 (1989) 317 -326. Edwards, W. , “How to Use Multiattribute Utility Measurement for Social Decision Making”, IEEE Transactions on Systems, Man, and Cybernetics 7 (1977) 326 -340. Ojanen, O. , Makkonen, S. and Salo, A. , “A Multi-Criteria Framework for the Selection of Risk Analysis Methods at Energy Utilities”, International Journal of Risk Assessment and Management (to appear). Salo, A. ja R. P. Hämäläinen, "Preference Assessment by Imprecise Ratio Statements”, Operations Research 40 (1992) 1053 -1061. Salo, A. and Hämäläinen, R. P. , “Preference Ratios in Multiattribute Evaluation (PRIME) - Elicitation and Decision Procedures under Incomplete Information”, IEEE Transactions on Systems, Man, and Cybernetics 31 (2001) 533 -545. Salo, A. and Liesiö, J. , “A Case Study in Participatory Priority-Setting for a Scandinavian Research Programme”, submitted manuscript. Salo, A. and Punkka, A. , “Rank Inclusion in Criteria Hierarchies”, European Journal of Operations Research (to appear). Stillwell, W. G. , Seaver, D. A. and Edwards, W. , “A Comparison of Weight Approximation Techniques in Multiattribute Utility Decision Making”, Organizational Behavior and Human Performance 28 (1981) 62 -77. 25
- Slides: 25