Testing satistical hipotheses in fuzzy environment
Abstract
In traditional statistics all parameters of the mathematical model and possible observations should be well defined. Sometimes such assumption appears too rigid for the real-life problems, especially while dealing with linguistic data or imprecise requirements. To relax this rigidity fuzzy methods are incorporated into statistics. We review hitherto existing achievements in testing statistical hypotheses in fuzzy environment, point out their advantages or disadvantages and practical problems. We propose also a formalization of that decision problem and indicate the directions of further investigations in order to construct a more general theory.
Full text
Mathware & Soft Computing 4 (1997) 203-217 Testing Statistical Hypotheses in Fuzzy Environment Przemys law Grzegorzewski, Olgierd Hryniewicz Systems Research Institute, Polish Academy of Sciences, l ul. Newelska 6, 01-447 Warsaw, Poland, e-mail: {pgrzeg, hryniewi}&ibspan.waw.pl Abstract In traditional statistics all parameters of the mathematical model and possible observations should be well defined. Sometimes such assumption appears too rigid for the real-life problems, especially while dealing with linguistic data or imprecise requirements. To relax this rigidity fuzzy methods are incorporated into statistics. We review hitherto existing achievements in testing statistical hypotheses in fuzzy environment, point out their advantages or disadvantages and practical problems. We propose also a formalization of that decision problem and indicate the directions of further investigations in order to construct a more general theory. Key words: hypothesis testing, fuzzy sets, fuzzy data, fuzzy hypothesis, the Neyman-Pearson lemma, the Bayesian approach. 1 Introduction At the first stage of investigations concerning different kind of phenomena (e.g., biological, technical, physical, and social) initial hypotheses relating to these phenomena are often formulated. Then, during the second experimental stage, facts which either confirm or falsify these hypotheses are collected and analyzed. When the considered hypothesis is of a probabilistic nature, i.e., the phenomenon under investigation is described by a probabilistic model, the methods of mathematical statistics are used. These methods, called statistical tests, let us specify such events which are almost improbable when the considered hypothesis is true. Observations of such events indicate that the considered hypothesis may be not true, and therefore, should be rejected. In traditional statistics all parameters of the mathematical model and the observed experimental data should be well defined. However, the complexity of investigated phenomena make the underlying models inadequate to the observed reality. In such situations the traditional models are valid only under some additional assumptions that might be not fulfilled. We face such a situation when our experimental data are of a linguistic type. 203
204 P. Grzegorzewski & O. Hryniewicz The problems stated above have motivated many researchers to enrich the traditional approach to hypothesis testing by introducing fuzzy models. These fuzzy models were proposed in order to describe uncertainties both in the experimental data and in the investigated hypotheses. Diversity of approaches indicates that we are yet in the initial stage, and the commonly accepted methodology hasn’t been worked out. Therefore, it seems to be useful to summarize the already obtained results in order to show their weak points. This is the aim of Sections 4, 5, 6 and 7 of this paper. Such an analysis can also show the directions of further investigations in this area (Sec. 8). In Section 2 we review some of the basic notions of the traditional approach to hypotheses testing. Section 3 provides a formalization of the problem in order to make more visible the differences between the traditional and the new approaches. 2 Elements of the traditional theory of testing statistical hypotheses Assume that the investigated phenomenon is described by a probability distribution Pθwhich belongs to a family of distributions P={Pθ:θ∈Θ}. We consider the null hypothesis H:θ∈ΘHconcerning the parameter θ, with the alternative hypothesis K:θ∈ΘK, where ΘHand ΘKare subsets of Θ such that ΘH∪ΘK= Θ and ΘH∩ΘK= Ø. We assume that if θwere known one would also know whether or not the hypothesis is true. If ΘH(ΘK) contains only a single element, then the considered hypothesis is called ”simple”, otherwise is said to be ”composite ”. In the experiment we observe a random variable X= (X1, X2, . . . , Xn), and this observation can lead to one of two possible decisions: either d0(to accept H, and to reject K), or d1(to accept K, and to reject H). Traditionally we denote d0 by zero, and d1by one. Hence, a decision rule (a statistical test) can be defined as a function ϕ:X → {0,1}, where Xis a sample space. Such a test is also called a non-randomized statistical test, contrary to a randomized statistical test eϕ:X → [0,1] which uses an additional random mechanism (independent from X) to undertake the decision. Each statistical test divides the sample space Xinto two exclusive subsets: {x∈ X :ϕ(x) = 0}- the set of the acceptance of H, and K={x∈ X :ϕ(x) = 1}- the set of the rejection of H(the acceptance of K) which is also called ”a critical region”. In practice, we compute a certain test statistic T(X) (i.e. a function of the observations), then we find a critical region K, and finally we reject the considered hypothesis if T(X)∈ K or accept it otherwise. Our decision can be either right or we can commit one of two possible errors: to reject Hwhen it is true (type I error), or to accept Hwhen it is false (type II error). The optimal test would minimize the probabilities of both errors simultaneously, but such a test doesn’t exist. Therefore, in practice we set an upper limit for the probability of type I error: Eθϕ(x)≤δ,∀θ∈ΘHwhere δis usually a small number (e.g. 0.01, or 0.05) and then we select a test which minimize the probability of type II error. In this case we also say that our test maximizes its power, i.e. the probability of the rejection of Hwhen it is false. The power of test is of interest
Testing Statistical Hypotheses in Fuzzy Environment 205 mainly while comparing different tests in order to find an optimal one. We say that the test ϕis more powerful than the test ψif Eθϕ(x)≥Eθψ(x),∀θ∈ΘK Eθϕ(x)> Eθψ(x),for a certain θ∈ΘK. The test which is more powerful than all other tests is called ”uniformly most powerful”. The approach to the problem of statistical testing described above was established by Neyman and Pearson. There exist also other approaches, e.g. Bayesian and minimax. Let L: Θ × {d0, d1}→Rbe a loss function. We interpret its value L(θ, di) as our loss when the true value of the parameter is θand we undertake the decision di. The expected value of the loss function R(θ, ϕ) = EθL(θ, ϕ(x)), θ ∈ΘH is called a risk. In the Bayesian approach a prior probability distribution π(θ) is imposed on the set Θ. The prior distribution π(θ) describes initial subjective opinions (a prior information) about possible values of θ. Knowing the prior distribution we can calculate the Bayes risk r(π, ϕ) = Z Θ R(θ, ϕ)π(dθ). For a given prior distribution an optimal Bayes test ϕis that one which minimizes the Bayes risk, i.e. r(π, ϕ)≤r(π, φ),∀φ∈Φ, where Φ is a set of possible tests. Unfortunately, in order to apply this method it is necessary to assume that θis a random variable and that its distribution π(θ) is known. This assumption is usually not warranted in applications. If no prior information regarding θis available one might consider the maximum of the risk function to be the most important feature. Of two tests the one with a smaller maximum is then preferable, and the optimal test ϕ0, called a minimax test, is that which minimizes the maximal risk,i.e. sup θ∈Θ R(θ, ϕ0) = inf ϕ∈Φsup θ∈Θ R(θ, ϕ). Since this maximum represents the worst average loss that can result from the use of a given procedure, the minimax solution is one that gives the greatest protection against large losses. For more details concerning the traditional theory of testing statistical hypotheses we refer the reader to [7, 22]. 3 A formal description of the hypotheses testing problem The notions defined in the previous section can be used for a formal description of the hypotheses testing problem.
206 P. Grzegorzewski & O. Hryniewicz Definition 1 Testing hypotheses is a decision problem described by an ordered 7- tuple (X,P,H,D,Φ,U,W), where: Xis a sample space (possible states), Pis a family of probability distributions, His a set of hypotheses, Dis a set of decisions, Φis a set of tests, Uis a utility function, and Wis a preference function on the set of tests Φ. The utility function Umay be interpreted (depending on the applied approach) as a confidence level, the loss function or the risk, whereas the preference function Wmay be interpreted as the power of test, the Bayes risk, or the maximal risk. Both Uand Wwe call sometimes requirements. As it has been already stated in Introduction, the traditional approach might not be sufficient in the real life. A generalization of the classical theory that takes into account non-random uncertainty and lack of precision is really needed in practice. A fuzzy approach could be useful to describe all imprecision by ”fuzzyfying” all elements of the 7-tuple which describes the decision problem in the presence of non-random uncertainties. However, it seems reasonable to restrict ourselves only to fuzzyfication of the space of possible states X, and of the set of hypotheses H. In the case of fuzzy states we can deal with imprecise observations such as: ”is close to 5”, ”somewhere between 6 and 8”, ”a lot of”, etc. In the case of fuzzy hypotheses we can formulate them, for example, in a form ”H:θis close to 5”, etc. Fuzzy observations may lead to fuzzy decisions Dand fuzzy tests Φ. The lack of precision can be also taken into account by the fuzzyfication of Uor W. 4 Testing hypotheses with fuzzy data Testing hypotheses with fuzzy data is discussed by Casals, Gil, and Gil [3, 4, 5, 8]. The starting point of their papers comes from the paper of Tanaka,Okuda, and Asai [31], who introduced the notion of the fuzzy information system A={(Ai, µi) : i∈I}, i.e., a set of fuzzy observations which fulfill additionally the orthogonality constraint: X i∈I µi(x) = 1,∀x∈S, where Ai⊂S⊆ R,Iis a set of indexes, and µi:S→[0,1] is a membership function. So we have X=An. In their papers, Casals,Gil, and Gil used Zadeh’s concept of the probability of fuzzy event e P(A1, . . . , An) = Z Sn µ1(x1). . . µn(xn)dP(x1, . . . , xn), where Pis a probability distribution on Sn. For the model given above, Casals, Gil, and Gil [3] proposed a new version of the Neyman-Pearson Lemma which supplies the uniformly most powerful test for
Testing Statistical Hypotheses in Fuzzy Environment 207 simple hypotheses H:θ=θ0,K:θ=θ1 ϕ(X1, . . . , Xn) = 1,when f P1(X1, . . . , Xn)> c0f P0(X1, . . . , Xn) 0,otherwise for a given significance level δ. Casals, Gil, and Gil proposed also a method of constructing the optimal Bayesian and minimax tests. The approach proposed by Casals, Gil, and Gil has, however, some weak points. First, a finite set of possible observations has to be specified in advance. This precludes the usage of their test in the cases of observations that are only countable or even uncountable. Secondly, the orthogonality constraint has to be fulfilled. That condition is rather of a technical nature and is not natural for practical applications. Moreover, it precludes the existence of two observations which are described simultaneously by nondecreasing or nonincreasing membership function (such as, for example, ”rather greater than 50”). In order to weaken the orthogonality constraint in the fuzzy information system Casals, Gil, and Gil [3] consider, so called, almost surely scale invariant tests. They conclude that given any set consisting of a finite number of fuzzy events we may construct a corresponding fuzzy information system leading to the same inference. More precisely, for any test ϕbased on fuzzy observations O={(Oi, νi) : i= 1,2, . . . , m}there exist a corresponding fuzzy information system A={(Ai, µi) : i= 1,2, . . . , m}and an optimal almost surely scale invariant test ϕ0based on A, for which probability of type I and type II error coincides with the corresponding probabilities of ϕ.O and Aare connected in the following way: ν1(x1). . . νm(xm) = Cµ1(x1). . . µm(xm) for almost all (x1, . . . xm)∈Sm, where Cis a given positive constant. Unfortunately the assertion that we can find a fuzzy information system Acorresponding to any space of fuzzy observations Osatisfying the equation given above is not true. A space Owhich includes at least two observations with nondecreasing or nonincreasing membership function is a simple counterexample. We can also prove a more general lemma: Lemma 1 Let Sbe a measurable space with a measure λ. Let O={(Oi, νi) : i= 1,2, . . . , m}denote a set of fuzzy observations, i.e., Oi⊂Sand νi:S→[0,1]. If there exist at least two observations (Oj, νj)and (Ok, νk)in Osuch that: 1. supp(νj)÷supp(νk)⊇U, where U⊂S,λ(U)>0and ÷is a symmetric difference of sets; 2. νj(x) = νk(x) = 1 for x∈V, where V⊂S,λ(V)>0; then it is not possible to construct the optimal almost surely scale invariant test corresponding to O.
208 P. Grzegorzewski & O. Hryniewicz Proof. Without loss of generality we may assume that m= 2. Let us take two fuzzy observations (O1, ν1), (O2, ν2) satisfying conditions 1), 2) of the lemma . Then we can take two points a∈Uand b∈Vsuch that ν1(a) = α, 0< α ≤1 ν1(b) = 1, ν2(a) = 0, ν2(b) = 1. Suppose, on the contrary, that we can construct the optimal almost surely scale invariant test corresponding to any set of fuzzy observations O. Then we should have a corresponding fuzzy information system (A1, µ1) and (A2, µ2) a positive constant Csuch that: ν1(x1)ν2(x2) = Cµ1(x1)µ2(x2) for almost all (x1, x2)∈S2.(1) We thus get ν1(a)ν2(a) = 0 = Cµ1(a)µ2(a),(2) ν1(b)ν2(b) = 1 = Cµ1(b)µ2(b),(3) ν1(a)ν2(b) = α=Cµ1(a)µ2(b),(4) ν1(b)ν2(a) = 0 = Cµ1(b)µ2(a).(5) Since (A1, µ1) and (A2, µ2) satisfy the orthogonality constraint, we also have µ1(a) + µ2(a) = 1,(6) µ1(b) + µ2(b) = 1.(7) From (2) we get µ1(a) = 0 or µ2(a) = 0, but from (4) we conclude that µ1(a)6= 0 and µ2(a) = 0. Similarly, from (5) it follows that µ1(b) = 0 or µ2(a) = 0, but (3) gives µ1(b)6= 0 and µ2(a) = 0. Hence (6) shows that µ1(a) = 1. Thus from (4) we get µ2(b) = α/C. Therefore (3) leads us to the conclusion, that µ1(b) = 1/α. If α < 1 then µ1(b)>1 which is impossible. If α= 1 then µ1(b) = 1, but from (7) we get µ2(b) = 0 which is impossible as well. The same reasoning applies to any choice of points a∈Uand b∈V. Thus equation (1) is not fulfilled on the set U×V. Since λ(U)>0 and λ(V)>0 we have λ(U×V)>0. This contradicts the assertion that we can find a fuzzy information system leading to the almost surely scale invariant test corresponding to any set of observations and proves our lemma. The reformulation of the Neyman-Pearson Lemma by Casals et al., though interesting, does not bring the problem close to reality. It is a well-known fact that the application of tests based directly on the Neyman-Pearson Lemma is minimal, and is limited only to the case of simple hypotheses. This lemma is extremely important as the starting point for the construction of other tests such as uniformly most powerful tests for composite hypotheses (the Karlin-Rubin theorem). It seems, however, that the generalization given by Casals, Gil, and Gil doesn’t let
Testing Statistical Hypotheses in Fuzzy Environment 209 us to make such an extension for testing composite hypotheses with fuzzy data. Moreover, the application of the test proposed by Casals et al. leads to very important computational problems. Practically these problems preclude the test from the usage even in the case of moderate samples. To overcome these problems Casals and Gil [5] proposed a modification based on the Central Limit Theorem. However, we don’t know the accuracy of this approximation. Another problem arises when we analyze the conditions under which the test based on the Zadeh’s definition of the probability could be used. In this approach the probability distribution is known explicitly, and the probabilities of events are not fuzzy. Hence, we deal with the case of uncertain observations of intrinsically non-fuzzy events driven by a probabilistic distribution of a known type, which raises some doubts. Moreover, no specific features of the fuzzy set theory are used except the notion of the membership functions, and thus this model could be regarded as a classical one only with a specific probability measure. The problem of testing hypotheses with fuzzy data is also considered by Son, Song and Kim [29]. In many situations it is hard to find the uniformly most powerful test based on the Neyman-Pearson lemma or such a test does not even exist. To overcome these difficulties the locally most powerful test which maximizes the slope of the power function of the test at a certain desired point has been suggested. Son, Song and Kim present a reformulation of the generalized Neyman- Pearson lemma and propose a construction of the locally most powerful test for the fuzzy information system. Their paper has all weak points analogous to those discussed above. The reader may also have an impression that the considered problems could be solved on the basis of the traditional probability theory and the use of fuzzy sets is a little bit artificial. 5 Testing fuzzy hypotheses The problem of testing fuzzy hypotheses has been considered independently by many authors. In their papers, techniques specific to fuzzy sets are widely used, contrary to the results described in the previous section. Therefore, new problems, non-existent in the traditional statistical approach, arise. 5.1 Bayesian approach Delgado, Verdegay, and Vila [6] consider the Bayesian approach to the problem of testing fuzzy hypotheses. They use the Representation (Decomposition) Theorem and transform the original fuzzy hypothesis Hto the problem of crisp hypotheses testing on a family of α-cuts Hα={θ∈Θ : µH(θ)> α}, α ∈[0,1]; H=[ α αHα, where µH: Θ →[0,1] is a membership function of the hypothesis H.
210 P. Grzegorzewski & O. Hryniewicz As a result of this simple and natural transformation we arrive, however, at a fuzzy decision. To deal with this problem Delgado et al. proposed to consider the following quantity α0= sup{α:Hαis accepted}. The value of α0describes a maximal α-cut for which the hypothesis is accepted. In their paper Delgado et al. considered two types of hypotheses. First, one-sided hypotheses defined by a membership function µH(θ) = 0, θ < θ0 Γ(θ), θ0≤θ < θ1 1, θ ≥θ1 where Γ : R → [0,1] is a continuous non decreasing (non increasing) function such that Γ(θ0) = 0, Γ(θ1) = 1 (or Γ(θ0) = 1, Γ(θ1) = 0). Secondly, two-sided hypotheses defined by a membership function µH(θ) = Mθ0−θ a, θ ≤θ0 Mθ−θ0 a, θ > θ0 where a∈ R+,M:R → [0,1] is a non increasing smooth function such that M(0) = 1 and limx→∞ M(x) = 0. The solution proposed by Delgado et al. has the virtue of simplicity. But it has its shortcomings: disregarding alternative hypotheses and considering only situation with a single observation. Another approach was proposed by Saade and Schwarzlander [27]. They consider the problem of testing a crisp null hypothesis H0:θ=θ0, where θis the expected value of the Gaussian distribution, against a fuzzy alternative H1:θ= θ0+a, where ais a normal triangular fuzzy number. Similarly in Saade [26] a fuzzy null hypothesis H0:θ=θ0+b, where bis also a triangular fuzzy number, is considered. In both cases the authors arrive at decisions described by fuzzy numbers. To obtain a crisp decision they suggest to use either a ”total distance criterion” T(a) = 1 Z0 1 2(A0 α+A00 α)dα or ”utility ranking index” Fu(a) = 1 Z0 ln (A0 αA00 α)dα where A0 α= inf{x∈ R :µa(x)≥α},A00 α= sup{x∈ R :µa(x)≥α},α∈[0,1]. In order to compare their test with a crisp minimax test Saade and Schwarzlander introduced as a preference function a ”fuzzy Bayes risk” defined as follows: FBR =Z Z0 [P0c00λ0R+P1c01λ1R]dR +Z Z1 [P0c10λ0R+P1c11λ1R]dR
Testing Statistical Hypotheses in Fuzzy Environment 211 where P0and P1are the prior probabilities of H0and H1, respectively; cij ,i, j = 0, describe the loss when the i-th hypothesis is accepted when the j-th is really true; Zi,i= 0,1, is a subset of the space of states for which the acceptance of the i-th hypothesis is preferred; λiR are so called fuzzy likelihood functions calculated for this test for given fuzzy numbers aand b(i.e. λiR is a fuzzy number induced from the fuzzy numbers aor b, respectively, by the map which is the conditional probability density function of the data under hypothesis Hi). The results of Saade and Schwarzlander were obtained only for a specific case: under rather restrictive assumptions of one observation, triangular fuzzy numbers, and the Gaussian distribution. Moreover, the authors didn’t analyze the choice of the appropriate index to compare two fuzzy numbers. However, their results are valuable for at least two reasons: both null and alternative hypotheses are considered and the obtained results were applied in practical situations [28]. Both preceding methods have all advantages and disadvantages of the Bayesian approach. Its weakest point is the necessity to know precisely a prior distribution of the considered parameter. But it is not our purpose to discuss here the principles of Bayesian analysis, which is a problem rather of a philosophical nature. We just want to point out only one strange feature of the methods presented above. Both Delgado et al. and Saade et al. assume that the prior distribution does not depend on the level of α-cuts, so it is constant. This condition is, in our opinion, questionable. Because for crisp observations and precisely defined prior distribution, using well-known statistical methods one would get well-defined and unique posterior distribution. Thus when we have precise prior information, testing fuzzy (i.e. imprecise) hypotheses is useless. In practice, it seems more natural to consider situations with fuzzy prior information, and then testing fuzzy hypotheses. Such approach to Bayes estimation and reliability analysis has been adapted by Hryniewicz [11, 14] and Nagata [23]. 5.2 Classical approach Classical, i.e. Neyman-Pearson, approach to the hypotheses testing was considered in the paper of Watanabe and Imaizumi [33]. They propose to transform a fuzzy problem into a set of crisp problems defined for α-cuts. For a given α-cut a classical test problem is solved. Then the results are aggregated in a form of a fuzzy decision Rrepresented by a membership function χR(0) = (sup {rδ(0)=0} µH(θ),if {θ:rδ(θ) = 0} 6=∅ 0,if {θ:rδ(θ) = 0}=∅ χR(1) = 1 −χR(0) where rδ(θ) = 0,if T(θ)/∈ K(δ, θ) 1,if T(θ)∈ K(δ, θ), T(θ) is a test statistics, and K(δ, θ) is a critical region for a given significance level δ.