Abstract for The Valuation of Choices in The Logic of Decision (154 words) Here is an example of an easy choice. When you visit a restaurant, would you prefer to have an entree selected from the menu by a lottery, or would you prefer to choose your dinner for yourself? While most individuals would have no hesitation identifying the opportunity to choose for oneself as preferable, most theories of rational decision are incapable of even posing the question. In this paper I will explain how within one influential system of analyzing decisions under risk, we may represent decision problems themselves as events which stand in relations of relative preference to other events. First I will show how to construct an algebra of choices from the algebra of events. Then I will show that the algebra of choices can be interpreted as including the original algebra of events. To conclude, I will compare the choice algebras constructed here to the representation of decision problems proposed by other authors. 1 The Valuation of Choices in The Logic of Decision Ken Presting presting@mindspring.com Here is an example of an easy choice. When you visit a restaurant, would you prefer to have an entree selected from the menu by a lottery, or would you prefer to choose your dinner for yourself? While most individuals would have no hesitation identifying the opportunity to choose for oneself as preferable, most theories of rational decision are incapable of even posing the question. This is perhaps understandable, since the problem traditionally1 called "decision making under risk" has not included an analysis of the value of choosing for oneself. The fact that in some situations one may choose for oneself has been represented only in the context of the discussion, not in the formal structure of the theory. By contrast, the theory of two-person zero-sum games has given explicit attention to the value of being in a game. Once a game has been "solved," that is, once an optimal strategy has been identified, then the value of the game is the expected utility of playing the optimal strategy. In this paper I will explain how within one influential system of analyzing decisions under risk, we may represent decision problems themselves as events which stand in relations of relative preference to other events. I will use Richard Jeffrey's "The 2 Logic of Decision" as my starting point, primarily because the abstract mathematical foundation of that system lends itself to extensions. Two basic concepts of Jeffrey's system are a Boolean algebra of events, and a preference relation defined over those events. First I will show how to construct an algebra of choices from the algebra of events. The algebra of choices represents all the possible decision problems which can be constructed from the given algebra of events. Then I will show that the algebra of choices can be interpreted as including the original algebra of events. Next, I will show that the preference relation may be naturally extended to compare the values of choice events to the simple events. To conclude, I will compare the choice algebras constructed here to the representation of decision problems proposed by other authors. Another useful feature is that the Desirability function of Jeffrey’s system effectively ranks the values of propositions in a one-dimensional continuous interval, like an interval in the real line. The values of propositions form equivalence classes, in which each point in a real interval is associated to a mutually indifferent set of propositions. Decision problems will also be mapped to equivalence classes, some of which correspond to real numbers. However, it will appear that certain cases of problems with infinitely many options correspond not to a specific real number, but to an open interval of reals. 3 Choice Events The event of having a choice is represented by a set. This "option set" set may be infinite, but may not be empty. The set represents a “menu”, and when a decisionmaking subject is presented with a menu, a subset of the menu items are selected. This "chosen subset" may be a singleton, and is also permitted to be empty 2. Formally, we begin with a complete and atom-free Boolean algebra of events, denoted by ℰ. Our option sets will be all the subsets of ℰ which exclude the null event “F”; in other words, the power set P(ℰ-{F}). Call the collection of option sets ℭ. We assume that the impossible event “F” can never be chosen. Note that in The Logic of Decision, “F” is excluded from the domain of the preference relation. The choice algebra ℭ is again a complete Boolean algebra, under the set operations of union, intersection, and complementation. An element in the choice algebra represents the opportunity to choose from among the events in the option set. Considerable work has been done (primarily by economists) in exploring the properties of choices represented by such menus. When a collection of menus is defined, a function which maps each menu to its chosen subset is called a "choice function." The question typically addressed by economists has been, "When does a choice function define a binary preference relation?" In this paper, I will be working in the other direction. We will assume that a preference relation exists, and use that relation to define a choice function. 4 The binary preference relation on the algebra of propositions ℰ is denoted by the symbol “≽” 3. Naturally, every subset of an option set is also a member of the choice algebra, so the choice functions have both a domain and range within the choice algebra ℭ. We can define a choice function ƒ from a preference relation by: [1] ƒ: ℭ → ℭ : ƒ(C) = {c ∈ C | x ∈ C ⇒ c ≽ x } In other words, given an option set C from the choice algebra ℭ, the chosen subset of C includes any elements of C which are preferable to all other element of C. Such elements are called “maximal” or “undominated”. Here are some useful observations: 1. If the chosen subset ƒ(C) has more than one member, then each member of ƒ(C) is indifferent to all the others4. 2. If C is finite, then ƒ(C) will never be empty. At least one element of C will be maximal. 3. If C is a singleton, then it sole member is itself maximal, and ƒ(C)=C. 4. If C is infinite, then C may contain a sequence of increasingly preferable members, and not contain a limit of that sequence. In such a case, no members are maximal and ƒ(C) will be the empty set. While the cases in (4) must be handled separately, we will see that they do not pose a difficulty. 5 Open Choices When an option set does not contain any undominated element, I will the decision problem "open”. In an open decision there are no choosable elements, and the chosen subset is empty. Even if an agent made a (perhaps arbitrary) selection of a suboptimal element, there is always a better option for each potential choice. Another twist, which does not work, is to choose all the elements whose value is above a certain threshold. But such a chosen subset would contain elements which are not indifferent. Consider the proposition which is the disjunction of such a subset. Then we can choose one of the dispreferred disjuncts, and form a new proposition which lacks just that dispreferred disjunct. Now the reduced proposition will be preferred to the original, because of the averaging property5. Thus we can construct a sequence of increasingly preferable chosen subsets, and again to select one rather than another is arbitrary. Although an open decision has no "solution" in the ideal sense 6 the open decisions may be assigned an unambiguous value. Consider first the case of an option set which is infinite but has an upper bound. An example using real numbers would be: O1 = { 1, 1.5, 1.75, 1.875, ... } The option set O1 has no maximal element, but at the same time, no element of O1 is as desirable as a proposition with value equal to two. Here are two observations regarding the set O1: 1. For any p in E st. p < 2, there is some o in O1 s.t. o > p 6 2. For and p in E st. p >= 2, there is no o in O1 st. o > p To have the opportunity to choose from among the elements of O1 must be preferable to simply getting any element of E with is strictly less desirable than two. But simply getting a proposition of desirability equal to two is preferable to any possible outcome of the decision problem posed by O1. Next, consider an option set containing an infinite sequence of elements with consistently increasing value. An example using real numbers would be: O2 = { 1, 3, 5, 7, ... } Again, the option set O2 has no maximal element. However, in this case there is no upper bound to the value of elements in O2. The relevant property of O2 is: 3. For any p in E, there is some o' in O2 st. o' > p That is to say, the opportunity to choose an element from O2 must be preferable to simply getting any proposition in E. Open decisions which are unbounded become maximal elements in the choice algebra. All these unbounded decision problems are in a single equivalence class, and are indifferent to each other. Preference for Choice Events Now we will formally extend the preference relation defined over the event algebra ℰ, to the comparison of option sets in ℭ. Take an arbitrary option set C, and consider the events in its chosen subset, ƒ(C). Since the elements of this subset are members of the complete Boolean algebra ℰ, the disjunction of all the elements in ƒ(C) is 7 also an element of ℰ. If the chosen subset of C is not empty, we will give the choice event C a position in the preference relation by setting it equal in preference to the disjunction of its chosen subset: [2] If ƒ(C) ≠ ∅, then C ≈ sup( ƒ(C) ). In this expression, “sup()” refers to the Boolean operation of the event algebra ℰ. The sup() or “supremum” corresponds to the potentially infinite disjunction of elements from the chosen subset ƒ(C)7. Now let us consider cases when ƒ(C ) is the empty set. These cases occur only when C does not contain any maximal element. This may occur in two ways. Either there is no upper bound for the preference of elements in C, or, there is an upper bound but not within C itself. If there is no upper bound for the preference of elements in C, then we will define the preference of C by: [3] If ∄e ∈ ℰ ∍ ∀c ∈ C, e ≽ c, then for any D ∈ ℭ, C ≽ D That is to say, an unbounded open decision problem is weakly preferred to any other choice. The unbounded choices are maximal, when the preference relation is extended to the choice algebra. The last case to consider is the case of a bounded option set which does not contain its limiting value. If C has any upper bound e ∈ ℰ, then by the continuity assumption8 for preference in ℰ, there is a unique equivalence class defined by the least upper bound LUB(C) = [e0] such that e0 is preferable to every element in C, and no other 8 element is lesser than e0 while still being an upper bound of C. Now we may define C as strongly preferable to each of its elements, but dis-preferred to its least upper bound: [4] If LUB(C) = [e0], and no e ∈ [e0] is also in C, then {e0} ≻ C and ∀c ∈ C, C ≻ {c} Taking the transitive closure completes the extension of the preference relation to CA. From these definitions some immediate consequences are: THEOREM I: Preference in ℭ is transitive, asymmetric, and connected. THEOREM 2: Preference in ℭ is everywhere continuous but nowhere dense. THEOREM 3: Preference in ℭ is homomorphic to the order relation among rightterminated open and closed half-lines of the real numbers, in terms of their right endpoints. 9 Option Sets and Matrix Representation Now let's compare the option set formalism to the decision matrix scheme which is more familiar in the philosophical literature on decision theory. We will see that the two approaches are equivalent. There are many versions of matrix representation under discussion in recent books and articles, but here I will consider only the system of James Joyce's Foundations of Causal Decision Theory9. For Joyce, a decision problem is a fourtuple <Ω,A,S,O>, where Ω is a Boolean sigma-algebra, Ai and Sj are partitions of the act-space and state-space, and Oij is the array of outcomes which obtain when act Ai is selected in state of nature Sj The matrix description translates directly into a option set. The elements of the option set are the "rows" of the decision matrix. More precisely, suppose that I and J are index sets, respectively, for the act partition Ai and the state partition Sj. The Cartesian product I x J is then the index set for the outcomes Oij. Now for any four-tuple <Ω,A,S,O> we define a option set by C = { Ci, i ∈ I | Ci = Union(j ∈ J)( Ai & Sj & Oij ) } Joyce considers only cases where Ω is a sigma-algebra and the index sets are countable. However, in this paper we are discussing the case envisioned by Jeffrey, in which the Boolean algebra of propositions is complete. In the complete case, unions of any arbitrary cardinality are allowed. 10 Constructing a matrix from an option set can be equally trivial, provided we don’t expect too much structure. In general, a detailed partition of states of nature will not be available for an option set, nor will the set of options form a partition of available acts, as Joyce’s system requires. In the special cases where the elements of the option set are mutually disjoint, Joyce’s partition requirement will be satisfied. Then a single state of nature, the necessary proposition “T”, will suffice as sole member of the State-partition Si. The same proposition will also serve as every member of the outcome array Oij. Thus, the option set generates a one-column matrix-form decision problem. The formal structure is given by: D = < ℰ, C, {T}, {T} > For the more general case of choice sets which have non-disjoint elements, the Ω for a Joyce-type matrix cannot be taken directly from the proposition algebra ℰ. Instead, we must assume a family of “ethically neutral,” mutually incompatible propositions ℬ, equal in cardinality to ℰ and both probabilistically and causally independent of the propositions in ℰ. Intuitively, the propositions in the “partition set” ℬ might represent verbal performances such as, “the agent chooses e1” or “the agent chooses e2”. It is convenient to take the propositions of ℬ as indexed by the propositions in ℰ. Now, for any option set C = {ci}, the Boolean algebra for the matrixform representation is generated by the set: M = { m ∈ ℬ ∪ ℰ | mi = bci & ci } 11 That is to say, the propositions needed to frame the decision problem as a matrix are formed by taking the conjunction of each member of the option set C, with the appropriately indexed member of the independent partition set ℬ. Then we again have a single-column matrix to represent the decision problem, where Ω is the Boolean algebra generated by M: D = < Ω, M, {T}, {T} > Conclusion We have seen that decision problems may be equally well represented by the simple menus of economic theory, as by the more elaborate matrix representations of philosophical decision theory. We have also seen that decision problems may have a value, even when they lack a solution. 1 For a classic exposition, see R. Duncan Luce and Howard Raiffa, Games and Decisions (New York: Wiley) 1957. 2 The formalism of decision sets and choice functions goes back at least to 1938. cf. Paul Samuelson, “A Note on the Pure Theory of Consumer’s Behavior, “ Economica, N.S vol. 5 (1938). For a survey see: Amartya Sen, “Choice Functions and Revealed Preference,” Review of Economic Studies (1970) 12 3 The preference relation is assumed to satisfy the hypotheses of Bolker's existence and uniqueness theorems. Cf. R. C. Jeffrey, The Logic of Decision, (Chicago: University of Chicago Press) 1983. 4 Because preference is an averaging relation, each member of the chosen subset is also indifferent to any disjunction of other members. Intuitively, we might say that any lottery among some of these indifferent elements is indifferent to any other. 5 Cf. R. C. Jeffrey, op. cit., p. 146 6 There is an extensive literature (originating in the work of Herbert Simon) on “satisficing” – the selection of sub-optimal elements from decision problems. I will not discuss those concepts in this paper. 7 Cf. R. C. Jeffrey, op. cit., p. 148 8 Ibid. 9 James M. Joyce, The Foundations of Causal Decision Theory (Cambridge: Cambridge University Press), 1999