Abstract
The menu-dependent nature of regret minimization creates subtleties when it is applied to dynamic decision problems. It is not clear whether forgone opportunities should be included in the menu. We explain commonly observed behavioral patterns as minimizing regret when forgone opportunities are present. If forgone opportunities are included, we can characterize when a form of dynamic consistency is guaranteed.
Similar content being viewed by others
Notes
Siniscalchi considers a more general information structure where the information that the DM receives can depend on her actions in an unpublished version of his paper (Siniscalchi 2006).
References
Al-Najjar, N., & Weinstein, J. (2009). The ambiguity aversion literature: A critical assessment. Economics and Philosophy, 25, 249–284.
Allais, M. (1953). Le comportement de l’homme rationnel devant le risque: Critique des postulats et axiomes de l’ecole americaine. Econometrica, 21(4), 503–546.
Anscombe, F., & Aumann, R. (1963). A definition of subjective probability. Annals of Mathematical Statistics, 34, 199–205.
Campos, L. M. D., & Moral, S. (1995). Independence concepts for sets of probabilities. (pp. 108–115). http://arxiv.org/abs/1302.4940.
Cattaneo, M. E. G. V. (2007). Statistical decisions based directly on the likeihood function. Ph.D. thesis, ETH.
Chateauneuf, A., & Faro, J. (2009). Ambiguity through confidence functions. Journal of Mathematical Economics, 45, 535–558.
Cousa, I., Moral, S., & Walley, P. (1999). Examples of independence for imprecise probabilities. In Proceedings of the first international symposium on imprecise probabilities and their applications (ISIPTA ’99).
de Cooman, G. (2005). A behavioral model for vague probability assessments. Fuzzy Sets and Systems, 154(3), 305–358.
Dubois, D., & Prade, H. (1998). Possibility measures: qualitative and quantitative aspects. In D. M. Gabbay & P. Smets (Eds.), Quantified representation of uncertainty and imprecision, handbook of defeasible reasoning and uncertainty management systems (Vol. 1, pp. 169–226). Dordrecht: Kluwer.
Ellsberg, D. (1961). Risk, ambiguity, and the savage axioms. The Quarterly Journal of Economics, 75(4), 643–669.
Epstein, L. G., & Le Breton, M. (1993). Dynamically consistent beliefs must be Bayesian. Journal of Economic Theory, 61(1), 1–22.
Epstein, L. G., & Schneider, M. (2003). Recursive multiple-priors. Journal of Economic Theory, 113(1), 1–31.
Gärdenfors, P., & Sahlin, N. (1982). Unreliable probabilities, risk taking, and decision making. Synthese, 53, 361–386.
Gärdenfors, P., & Sahlin, N. (1983). Decision making with unreliable probabilities. British Journal of Mathematical and Statistical Psychology, 36, 240–251.
Gilboa, I., & Schmeidler, D. (1993). Updating ambiguous beliefs. Journal of Economic Theory, 59, 33–49.
Good, I. J. (1980). Some history of the hierarchical Bayesian methodology. In J. M. Bernardo, M. H. DeGroot, D. Lindley, & A. Smith (Eds.), Bayesian statistic I (pp. 489–504). Valencia: University Press.
Gul, F., & Pesendorfer, W. (2005). The revealed preference theory of changing tastes. Review of Economic Studies, 72(2), 429–448.
Halpern, J.Y., & Leung, S. (2012). Weighted sets of probabilities and minimax weighted expected regret: new approaches for representing uncertainty and making decisions. In Proceedings of the twenty-ninth conference on uncertainty in artificial intelligence (UAI 2012) (pp. 336–345).
Hammond, P. J. (1976). Changing tastes and coherent dynamic choice. The Review of Economic Studies, 43(1), 159–173.
Hanany, E., & Klibanoff, P. (2009). Updating ambiguity averse preferences. The B.E. Journal of. Theoretical Economics, 9(1) (Article 37)
Hayashi, T. (2009). Stopping with anticipated regret. Journal of Mathematical Economics, 45(7–8), 479–490.
Hayashi, T. (2011). Context dependence and consistency in dynamic choice under uncertainty: The case of anticipated regret. Theory and Decision, 70, 399–430.
Klibanoff, P., & Hanany, E. (2007). Updating preferences with multiple priors. Theoretical Economics, 2(3), 261–298.
Klibanoff, P., Marinacci, M., & Mukerji, S. (2005). A smooth model of decision making under ambiguity. Econometrica, 73(6), 1849–1892.
Kreps, D. M. (1988). Notes on the theory of choice. Boulder: Westview Press.
Levi, I. (1985). Imprecision and uncertainty in probability judgment. Philosophy of Science, 52, 390–406.
Maccheroni, F., Marinacci, M., & Rustichini, A. (2006a). Ambiguity aversion, robustness, and the variational representation of preferences. Econometrica, 74(6), 1447–1498.
Maccheroni, F., Marinacci, M., & Rustichini, A. (2006b). Dynamic variational preferences. Journal of Economic Theory, 128(1), 4–44.
Machina, M. J. (1989). Dynamic consistency and non-expected utility models of choice under uncertainty. Journal of Economic Literature, 27(4), 1622–68.
Nau, R. F. (1992). Indeterminate probabilities on finite sets. Annals of Statistics, 40(4), 1737–1767.
Rockafellar, R. T. (1970). Convex analysis. Princeton: Princeton University Press.
Savage, L. J. (1951). The theory of statistical decision. Journal of the American Statistical Association, 46, 55–67.
Savage, L. J. (1954). The foundations of statistics. New York: Wiley.
Siniscalchi, M. (2006). Dynamic choice under ambiguity. Discussion papers, Northwestern University, Center for Mathematical Studies in Economics and Management Science.
Siniscalchi, M. (2011). Dynamic choice under ambiguity. Theoretical Economics, 6(3), 379–421.
Stoye, J. (2011). Axioms for minimax regret choice correspondences. Journal of Economic Theory, 146(6), 2226–2251.
Strotz, R. H. (1955). Myopia and inconsistency in dynamic utility maximization. The Review of Economic Studies, 23(3), 165–180.
Walley, P. (1991). Statistical reasoning with imprecise probabilities, monographs on statistics and applied probability (Vol. 42). London: Chapman and Hall.
Walley, P. (1997). Statistical inferences based on a second-ordr possibility distribution. International Journal of General Systems, 26(4), 337–383.
Zadeh, L. A. (1978). Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems, 1, 3–28.
Author information
Authors and Affiliations
Corresponding author
Additional information
Work supported in part by NSF Grants IIS-0812045, IIS-0911036, and CCF-1214844, by AFOSR Grants FA9550-08-1-0438, FA9550-09-1-0266, and FA9550-12-1-0040, and by ARO Grant W911NF-09-1-0281. We thank an anonymous reviewer and Spyros Galanis for useful comments.
Appendices
Proof of Theorem 4.1
We restate the theorem (and elsewhere in the appendix) for the reader’s convenience.
Theorem 4.1. For a dynamic decision problem D, if and \(\mu (h)=M\) for some fixed menu M, then there will be no preference reversals in D.
Proof
Before proving the result, we need some definitions. Say that an information set I refines an information set \(I'\) if, for all \(h \in I\), some prefix \(h'\) of h is in \(I'\). Suppose that there is a history h such that \(f, g \in M_h\) and \(I(h) = I\). Let fIg denote the plan that agrees with f at all histories \(h'\) such that \(I(h')\) refines I and agrees with g otherwise. As we now show, fIg gives the same outcome as f on states in \(E = E(h)\) and the same outcome as g on states in \(E^c\); moreover, \(fIg \in M_h\).
Suppose that \(s(h) = s\) and that \(s \in E\). Since \(E(h) = E\), there exists a history \(h' \in I(h)\) such that \(s(h') = s'\) and \(R(h') = R(h)\). Since \(f, g \in M_h\), there must exist some k such that \(f^k(\langle s\rangle ) = g^k(\langle s\rangle ) = h\) [where, as usual, \(f^0(\langle s\rangle ) = \langle s\rangle \) and for \(k' \ge 1\), \(f^{k'}(\langle s\rangle ) = f(f^{k'-1}(\langle s\rangle ))\)]. We claim that for all \(k' \le k\), \(f^{k'}(\langle s'\rangle ) = g^{k'}(\langle s'\rangle )\), and \(f^{k'}(\langle s'\rangle )\) is in the same information set as \(f^{k'}(\langle s\rangle )\). The proof is by induction on \(k'\). If \(k' = 0\), the result follows from the observation that since \(\langle s\rangle \) is a prefix of h, there must be some prefix of \(h'\) in \(I(\langle s\rangle )\). For the inductive step, suppose that \(k' \ge 1\). We must have \(f^{k'}(\langle s\rangle ) = g^{k'}(\langle s\rangle )\) (otherwise g would not be in \(M_h\)). Since \(g^{k'-1}(\langle s\rangle ) = f^{k'-1}(\langle s\rangle )\) and \(f^{k'-1}(\langle s'\rangle ) = g^{k'-1}(\langle s'\rangle )\) are in the same information set, by the inductive hypothesis, g must perform the same action at \(g^{k'-1}(\langle s\rangle )\) and \(g^{k'-1}(\langle s'\rangle )\), and must perform the same action at \(f^{k'-1}(\langle s\rangle )\) and \(f^{k'-1}(\langle s'\rangle )\). Since \(g^{k'}(\langle s\rangle )\) and \(f^{k'}(\langle s\rangle )\) are both prefixes of h, g and f perform the same action at \(f^{k'-1}(\langle s\rangle ) = g^{k'-1}(\langle s\rangle )\). It follows that f and g perform the same action at \(f^{k'-1}(\langle s'\rangle ) = g^{k'-1}(\langle s'\rangle )\), and so \(f^{k'}(\langle s'\rangle ) = g^{k'}(\langle s'\rangle )\). Thus, \(g^{k'}(\langle s'\rangle )\) must be a prefix of \(h'\), and so must be in the same information set as \(f^{k'}(\langle s\rangle )\). This completes the inductive proof.
Since \(f^k(\langle s'\rangle ) = g^k(\langle s'\rangle ) =h'\), it follows that \(f^k(\langle s'\rangle ) = (fIg)^k(\langle s'\rangle )\). Below I, all the information sets are refinements of I, so by definition, for \(k' \le k\), we must \(f^{k'}(\langle s'\rangle ) = (fIg)^{k'}(\langle s'\rangle )\). Thus, f and fIg give the same outcome for \(s'\), and hence all states in E. Note it follows that \((fIg)^k(\langle s\rangle ) = h\), so \(fIg \in M_h\).
For \(s' \notin E\) and all \(k'\), it cannot be the case that \(I((fIg)^{k'}(\langle s'\rangle ))\) is a refinement of I, since the first state in \((fIg)^{k'}(\langle s'\rangle ))\) is \(s'\), and no history in a refinement of I has a first state of \(s'\). Thus, \(fIg^{k'}(\langle s'\rangle ) = g^{k'}(\langle s'\rangle )\) for all \(k'\), so f and fIg give the same outcome for \(s'\), and hence all states in \(E^c\).
Returning to the proof of the proposition, suppose that \(f \in C_{\mu ,h}(M_h)\), \(h'\) is a history extending h, and \(f \in M_{h'}\). We want to show that \(f \in C_{\mu ,h'}(M_{h'})\). By perfect recall, \(E(h') \subseteq E(h)\). Suppose, by way of contradiction, that \(f \notin C_{\mu ,h'}(M_{h'})\). Since \(f \in C_{\mu ,h'}(M_{h'})\), we cannot have \(E(h') = E(h)\), so \(E(h') \subset E(h)\). Choose \(f' \in C_{\mu ,E(h')}(M_{h'})\) and \(g \in C_{\mu ,E(h')^c \cap E(h)}(M_{h'})\) (note that \(C_{\mu ,E(h')}(M_{h'}) \ne \emptyset \) and \(C_{\mu ,E(h')^c \cap E(h)}(M_{h'}) \ne \emptyset \) by Axiom 3). Since \(f', g \in M_{h'}\) (by Axiom 3), \(f'I(h')g\) is in \(M_{h'}\). Since \(f'I(h')g\) and \(f'\), when viewed as acts, agree on states in \(E(h')\), we must have \(f'I(h')g \in C_{\mu ,E(h')}(M_{h'})\) by Axiom 2. Similarly, since \(f'I(h')g\) and g, when viewed as acts, agree on states in \(E(h')^c \cap E(h) \), we must have \(f'I(h')g \in C_{\mu ,E(h')^c \cap E(h)}(M_{h'})\). Therefore, by Axiom 1, \(f'I(h')g \in C_{\mu ,h}(M_{h'})\). Also by Axiom 1, since \(f \notin C_{\mu ,h'}(M_{h'})\), we must have \(f \notin C_{\mu ,h}(M_{h'})\). By Axiom 4, this implies that \(f \notin C_{\mu ,h}(M_{h})\) (since \(M_{h'} \subseteq M_h\)), giving us the desired contradiction. \(\square \)
Proof of Theorem 4.3
Theorem 4.3. If \({\mathcal {P}}^+\) is a set of weighted distributions on \((S,\Sigma )\) such that \(C({\mathcal {P}}^+)\) is closed, then the following are equivalent:
-
(a)
For all decision problems D based on \((S,\Sigma )\) and all menus M in D, Axioms 1–4 hold for choice functions represented by \({\mathcal {P}}^+|^l E\) (resp., \({\mathcal {P}}^+|^p E)\).
-
(b)
For all decision problems D based on \((S,\Sigma )\), states \(s \in S\), and acts \(f\in M_{\langle s \rangle }\), the weighted regret of f with respect to \(M_{\langle s \rangle }\) and \({\mathcal {P}}^+\) is separable.
We actually prove the following stronger result.
Theorem 7.1
If \({\mathcal {P}}^+\) is a set of weighted distributions on \((S,\Sigma )\) such that \(C({\mathcal {P}}^+)\) is closed, then the following are equivalent:
-
(a)
For all decision problems D based on \((S,\Sigma )\), Axioms 1–4 hold for menus of the form \(M_{\langle s \rangle }\) for choice functions represented by \({\mathcal {P}}^+|^l E\) (resp., \({\mathcal {P}}^+|^p E)\).
-
(b)
For all decision problems D based on \((S,\Sigma )\) and all menus M in D, Axioms 1–4 hold for choice functions represented by \({\mathcal {P}}^+|^l E\) (resp., \({\mathcal {P}}^+|^p E)\).
-
(c)
For all decision problems D based on \((S,\Sigma )\), states \(s \in S\), and acts \(f\in M_{\langle s \rangle }\), the weighted regret of f with respect to \(M_{\langle s \rangle }\) and \({\mathcal {P}}^+\) is separable.
-
(d)
For all decision problems D based on \((S,\Sigma )\), menus M in D, and acts \(f\in M\), the weighted regret of f with respect to M and \({\mathcal {P}}^+\) is separable.
Proof
Fix an arbitrary state space S, measurable events \(E, F \subseteq S\), and a set \({\mathcal {P}}^+\) of weighted distributions on \((S,\Sigma )\). The fact that (b) implies (a) and (d) implies (c) follows immediately. Therefore, it remains to show that (a) implies (d) and that (c) implies (b).
Since the proof is identical for prior-by-prior updating (\(|^p\)) and for likelihood updating (\(|^l\)), we use | to denote the updating operator. That is, the proof can be read with | denoting \(|^p\), or with | denoting \(|^l\).
To show that (a) implies (d), we first show that Axiom 1 implies that for all decision problems D based on \((S,\Sigma )\), menu M in D, sets \({\mathcal {P}}^+\) of weighted probabilities, and acts \(f \in M\),
Suppose, by way of contradiction, that (1) does not hold. Then for some decision problem D based on \((S,\Sigma )\), measurable events \(E,F \subseteq S\), menu M in D, and act \(f\in M\), we have that
We define a new decision problem \(D'\) based on \((S,\Sigma )\). The idea is that in \(D'\), we will have a plan \(a_{f'}\) such that \(a_{f'} \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',E\cap F}}(M'')\) and \(a_{f'}\in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',E^c \cap F}}(M'')\) and \(a_{f'}\notin C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',F}}(M'')\) for some \(M''\subseteq M'\), where \(M'\) is the menu at the initial decision node for the DM.
We construct \(D'\) as follows. \(D'\) is a depth-two tree; that is, nature makes a single move, and then the DM makes a single move. At the first step, nature choose a state \(s \in F\). At the second step, the DM chooses from the set \(\{ a_g : g \in M \} \cup \{a_{f'}\}\) of actions. With a slight abuse of notation, we let \(a_g\) also denote the plan in \(T'\) that chooses the action \(a_g\) at the initial history \(\langle s\rangle \). Therefore, the initial menu in decision problem \(D'\) is \(M'=\{ a_g : g \in M \} \cup \{a_{f'}\}\).
The utilities for the actions/plans in \(D'\) are defined as follows. For actions \(\{a_g : g\in M\}\), the utility of \(a_g\) in state s is just the utility of the outcome resulting from applying plan g in state s in decision problem D. The action \(a_{f'}\) has utilities
For all states \(s\in F\), we have that \(u(a_{f'}(s)) \le \sup _{g\in M} u(g(s))\). As a result, for all states \(s\in F\), we have that
Since the regret of a plan in state s depends only on its payoff in s and the best payoff in s, it is not hard to see that the regrets of \(a_g\) with respect to \(M'\) is the same as the regret of g with respect to M. More precisely, for all \(g \in M\),
By definition of \(a_{f'}\), for each state \(s\in E\cap F\), we have \({\mathrm {reg}}_{M'}(a_{f'},s) = {\mathrm {reg}}_{M}^{{\mathcal {P}}^+|(E\cap F)}(f)\), and for each state \(s\in E^c \cap F\), we have \({\mathrm {reg}}_{M'}(a_{f'},s) = {\mathrm {reg}}_{M}^{{\mathcal {P}}^+|(E^c \cap F)}(f)\). Thus, for all \(\Pr \in {\mathcal {P}}\), if \(\Pr (E\cap F) \ne 0\), then \({\mathrm {reg}}_M^{\Pr |(E\cap F)}(f) = {\mathrm {reg}}_{M}^{{\mathcal {P}}^+|(E\cap F)}(f)\), and if \(\Pr (E^c\cap F) \ne 0\), then \({\mathrm {reg}}_M^{\Pr |(E^c\cap F)}(f) = {\mathrm {reg}}_{M}^{{\mathcal {P}}^+|(E^c\cap F)}(f)\). If for all \((\Pr ,\alpha ) \in {\mathcal {P}}^+|(E\cap F)\), \(\alpha \Pr ( E\cap F) = 0\), then \({\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E\cap F)}(a_{f'}) = {\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E\cap F)}(a_f) = 0\). Otherwise, since there is some measure in \({\mathcal {P}}^+|(E\cap F)\) that has weight 1, we must have \({\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E\cap F)}(a_{f'}) = {\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E\cap F)}(a_f)\). Similarly, \({\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E^c \cap F)}(a_{f'}) = {\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E^c \cap F)}(a_f)\). Thus,
Therefore, we have \(a_{f'} \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',E\cap F}}(\{a_{f'}, a_f\})\), \(a_{f'} \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',E^c \cap F}}(\{a_{f'}, a_f\})\), and \(a_{f'}\notin C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',F}}(\{a_{f'}, a_f\})\), violating Axiom 1.
By an analogous argument, we show that the opposite weak inequality,
is also implied by Axiom 1. Suppose, by way of contradiction, that (2) does not hold. Then for some decision problem D based on \((S,\Sigma )\), measurable events \(E,F \subseteq S\), menu M in D, and act \(f\in M\), we have that
We define a decision problem \(D'\) based on \((S,\Sigma )\) just as in the previous case. Specifically, we have that \({\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E\cap F)}(a_{f'}) = {\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E\cap F)}(a_f)\), and that \({\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E^c \cap F)}(a_{f'}) = {\mathrm {reg}}_{M'}^{{\mathcal {P}}^+|(E^c \cap F)}(a_f)\). The one difference from the previous case is that we now have
Therefore, we have \(a_{f} \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',E\cap F}}(\{a_{f'}, a_f\})\), \(a_{f} \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',E^c \cap F}}(\{a_{f'}, a_f\})\), and \(a_{f}\notin C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',F}}(\{a_{f'}, a_f\})\), violating Axiom 1.
To complete the proof that (a) implies (d), we show that Axiom 1 also implies that for all decision problems D based on \((S,\Sigma )\), menus M in D, sets \({\mathcal {P}}^+\) of weighted probabilities, and acts \(f \in M\), if \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+|(E\cap F)}(f) > 0 \), then
Suppose, by way of contradiction, that (3) does not hold. Then for some decision problem D based on \((S,\Sigma )\), events \(E,F\subseteq S\), menu M in D, and act \(f\in M\) such that \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+|(E\cap F)}(f) > 0 \) and
We now define a new decision problem \(D'\) based on \((S,\Sigma )\). The idea is that in \(D'\), we have a plan \(a_{f}\) such that \(a_{f} \notin C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M,E\cap F}}(M')\) but \(a_{f} \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M,F}}(M')\) for some \(M'\subseteq M\).
Construct \(D'\) exactly as before. That is, in the first step, nature chooses a state \(s\in S\), and in the second step, the DM chooses from the set of actions/plans \(M' = \{ a_g : g \in M \} \cup \{ a_{g'}\}\). For each \(g\in M\), define the actions \(a_g\) as before. We define a new action \(a_{g'}\) with utilities
It is almost immediate from the definition of \(a_{g'}\) that we have
However, we also have
Therefore, we have \(a_{f} \notin C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',E\cap F}}(\{a_{g'}, a_f\})\) but \(a_{f}\in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{{M',F}}(\{a_{g'}, a_f\})\), violating Axiom 1.
We next show that (c) implies (b). Specifically, we show that SEP for the initial menus of all decision problems D is sufficient to guarantee that Axioms 1–4 hold for menu M and all choice sets \(M' \subseteq M\). It is easy to check that Axioms 2–4 hold for MWER, so we need to check only Axiom 1.
Consider an arbitrary decision problem D, menu M in D, \(M'\subseteq M\), and a plan f in \(M'\). We construct a new decision problem \(D'\) such that the initial menu of \(D'\) is “equivalent” to M. Just as before, let \(D'\) be a two-stage decision problem where in the first stage, nature chooses \(s\in S\), and in the second stage, the DM chooses from the set \(M_0 = \{ a_g : g \in M\}\), where \(a_g\) is defined as before. Again, we associate each action \(a_g\) with the plan that chooses \(a_g\) in \(D'\). \(M_0\) is then “equivalent” to M in the sense that
Suppose that \(f \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{M,E\cap F}(M')\) and \(f \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{M,E^c \cap F}(M')\). This means that for all \(g \in M'\), we have \({\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E\cap F)}(a_f) \le {\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E\cap F)}(a_g)\) and \({\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E^c \cap F)}(a_f) \le {\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E^c \cap F)}(a_g)\). Therefore, we have
which means that \(f \in C^{{\mathrm {reg}},{\mathcal {P}}^+}_{M,F}(M')\), as required.
Next, consider an act \(g \in M'\) such that \(g \notin C^{{\mathrm {reg}},{\mathcal {P}}^+}_{M,E\cap F}(M')\). This means that \({\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E \cap F)}(a_f) < {\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E \cap F)}(a_g)\) and \({\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E^c \cap F)}(a_f) \le {\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E^c \cap F)}(a_g)\). Let \((\alpha _{\Pr ^*},\Pr ^*) \in C({\mathcal {P}}^+)\) be such that
Such a pair \((\alpha _{\Pr ^*},\Pr ^*)\) exists, since we have assumed that \(C({\mathcal {P}}^+)\) is closed. If \(\alpha _{\Pr ^*}{\Pr }^*(E\cap F) = 0\), then \({\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|F}(a_g) = \sup _{(\Pr ,\alpha ) \in {\mathcal {P}}^+} \alpha \left( \Pr (E^c \cap F) {\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E^c \cap F)}\right. \left. (a_g) \right) \). By separability, it must be the case that \({\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E \cap F)}(a_g) = 0\), contradicting our assumption that \(0 \le {\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E \cap F)}(a_f) < {\mathrm {reg}}_{M_0}^{{\mathcal {P}}^+|(E \cap F)}(a_g)\). Therefore, it must be that \(\alpha _{\Pr ^*}\Pr ^*(E\cap F) > 0\), and
which means that \(g \notin C^{{\mathrm {reg}},{\mathcal {P}}^+}_{M,F}(M')\). \(\square \)
Proof of Theorem 4.5
To prove Theorem 4.5, we need the following lemma.
Lemma 8.1
For all utility functions u, sets \({\mathcal {P}}^+\) of weighted probabilities, acts f, and menus M containing f, \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+}(f) = {\mathrm {reg}}_{M}^{C({\mathcal {P}}^+)}(f)\).
Proof
Simply observe that
by definition. \(\square \)
The next lemma uses an argument almost identical to one used in Lemma 7 of Halpern and Leung (2012).
Lemma 8.2
If \(C({\mathcal {P}}^+|^{\chi } F)\) is convex and q is a subprobability on F not in \(\overline{C({\mathcal {P}}^+|^{\chi }F)}\), then there exists a non-negative vector \(\theta \) such that for all \((\Pr ,\alpha ) \in {\mathcal {P}}^+|^{\chi }F\), we have
Proof
Given a set \({\mathcal {P}}^+\) of weighted probabilities, let \(C'({\mathcal {P}}^+) = \{ p : p \in {\mathbb {R}}^{|S|} \text { and } p \le \alpha \Pr \text { for some } (\Pr ,\alpha ) \in {\mathcal {P}}^+ \}\). Note that an element \(q \in C'({\mathcal {P}}^+)\) may not be a subprobability measure, since we do not require that \(q(s) \ge 0\). Since \(\overline{C'({\mathcal {P}}^+|^{\chi } F)}\) and \(\{{q}\}\) are closed, convex, and disjoint, and \(\{{q}\}\) is compact, the separating hyperplane theorem Rockafellar (1970) says that there exist \(\theta \in {\mathbb {R}}^{|S|}\) and \(c\in {\mathbb {R}}\) such that
Since \(\{ \alpha \Pr : (\Pr ,\alpha ) \in {\mathcal {P}}^+|^{\chi } F \} \subseteq \overline{C'({\mathcal {P}}^+|^{\chi } F)}\), we have that for all \((\Pr ,\alpha ) \in {\mathcal {P}}^+|^{\chi }F\),
Now we argue that it must be the case that \(\theta (s) \ge 0\) for all \(s\in F\). Suppose that \(\theta (s') < 0 \) for some \(s'\in F\). Define \(p^*\) by setting
Note that \(p^{*} \le \mathbf {0}\), since for all \(s\in S\), \(p^{*}(s) \le 0\). Therefore, \(p^{*} \in C'({\mathcal {P}}^+|^\chi F)\).
Our definition of \(p^{*}\) also ensures that \(\theta \cdot {p^{*}} = \sum _{s\in S} p^{*}(s) \theta (s) = p^{*}(s') \theta (s') = |c| \ge c \). This contradicts (4), which says that \(\theta \cdot {p} < c \text { for all } {p}\in C'({\mathcal {P}}^+|^{\chi } F)\). Thus, it must be the case that \(\theta (s) \ge 0\) for all \(s\in S\). \(\square \)
We are now ready to prove Theorem 4.5, which we restate here.
Theorem 4.5 If \(C({\mathcal {P}}^+)\) is closed and convex, then Axiom 1 holds for the family of choices \(C_{M}^{{\mathrm {reg}},{\mathcal {P}}^+ |^{\chi } E}\) if and only if \({\mathcal {P}}^+\) is \(\chi \)-rectangular.
We prove the two directions of implication in the theorem separately. Note that the proof that \(\chi \)-rectangularity implies Axiom 1 does not require \(C({\mathcal {P}}^+)\) to be convex.
Claim 8.3
If \({\mathcal {P}}^+\) is \(\chi \)-rectangular, then Axiom 1 holds for the family of choices \(C_{M}^{{\mathrm {reg}},{\mathcal {P}}^+ |^{\chi } E}\).
Proof
By Theorem 4.3, it suffices to show that SEP holds. For the first part of SEP, we must show that
Unwinding the definitions, (5) is equivalent to
The \(\sup \)s in this expression are taken on by some \((\Pr _1^*,\alpha _1^*),(\Pr _2^*,\alpha _2^*),(\Pr _3^*,\alpha _3^*) \in \overline{{\mathcal {P}}^+|^{\chi }F}\). By \(\chi \)-rectangularity, we have that for all \((\Pr _1,\alpha _1),(\Pr _2,\alpha _2),(\Pr _3,\alpha _3) \in {\mathcal {P}}^+|^{\chi }F\),
Thus, for all \(\epsilon > 0\),
Therefore,
as required.
It remains to show the opposite inequality in (5), namely, that
It suffices to note that the right-hand side is equal to
This completes the proof that (5) holds.
For the second part of SEP, suppose that \(\overline{{\mathcal {P}}}^+(E\cap F) > 0\) and \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+|^{\chi }(E \cap F)}(f) \ne 0\). If \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+|^{\chi }(E^c \cap F)}(f) = 0\) then, since \(\overline{{\mathcal {P}}}^+(E\cap F) > 0\), we have that \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+|^{\chi } F}(f) > 0 = \sup _{(\Pr ,\alpha ) \in {\mathcal {P}}^+|^{\chi }F} \alpha Pr(E^c \cap F) {\mathrm {reg}}_{M}^{{\mathcal {P}}^+|^{\chi }(E^c \cap F)}(f)\), as desired. Otherwise, by part (b) of \(\chi \)-rectangularity, for all \(\delta > 0\), there exists \((\Pr ,\alpha ) \in {\mathcal {P}}^+|^{\chi } F\) such that \(\alpha ( \delta \Pr (E \cap F) + \Pr (E^c \cap F)) > \sup _{(\Pr ',\alpha ') \in {\mathcal {P}}^+} \alpha ' \Pr '(E^c \cap F)\). Therefore, using the first part of SEP, we have
as required. \(\square \)
Claim 8.4
If \(C({\mathcal {P}}^+)\) is convex and Axiom 1 holds for the family of choices \(C_{M}^{{\mathrm {reg}},{\mathcal {P}}^+ |^{\chi } E}\), then \({\mathcal {P}}^+\) is \(\chi \)-rectangular.
Proof
Suppose that \(\chi \)-rectangularity does not hold. Then one of the three conditions of rectangularity must fail.
First suppose that it is (a); that is, for some \((\Pr _1,\alpha _1), (\Pr _2,\alpha _2), (\Pr _3,\alpha _3) \in {\mathcal {P}}^+\), we have \( \Pr _1( E\cap F) > 0 \) and \(\Pr _2(E^c \cap F) > 0\) and
Let \(p^* = \alpha _3 {\Pr }_3( E \cap F) \alpha _{1,E\cap F}^{\chi } {\Pr }_1| (E \cap F) + \alpha _3 {\Pr }_3(E^c\cap F) \alpha _{2,E^c \cap F}^{\chi } {\Pr }_2|(E^c\cap F)\). Since we have assumed that \(C({\mathcal {P}}^+)\) is convex, we have that \(C({\mathcal {P}}^+|^{\chi } F)\) is also convex. By Lemma 8.2, there exists a non-negative vector \(\theta \) such that for all \(\alpha \Pr \in \overline{C({\mathcal {P}}^+|^{\chi }F)}\), we have
We construct a decision problem D based on \((S,\Sigma )\). D has two stages: in the first stage, nature chooses a state \(s \in S\), but only states in \(F\subseteq S\) are chosen with positive probability, so when the DM plays, his beliefs are characterized by \({\mathcal {P}}^+|^{\chi }F\). In the second stage, the DM chooses an action from the set \(M = \{f,g\}\), with utilities defined as follows:
The act f will have regret precisely \(\theta (s)\) in state \(s\in S\). By Lemma 8.2,
violating SEP. By Theorem 4.3, Axiom 1 cannot hold.
Now suppose that condition (b) in rectangularity does not hold. That is, for some \(\delta > 0\), for all \((\alpha ,\Pr ) \in {\mathcal {P}}^+\), \(\alpha ( \delta \Pr (E \cap F) + \Pr (E^c \cap F) ) \le \sup _{(\Pr ',\alpha ') \in {\mathcal {P}}^+} \alpha ' \Pr '(E^c \cap F)\). We construct a decision problem D based on \((S,\Sigma )\). D has two stages: in the first stage, nature chooses a state \(s \in S\). In the second stage, the DM chooses an action from the set \(M = \{f,g\}\), with utilities defined as follows:
Then we have that \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+|^{\chi }(E \cap F)}(g) = \delta \) and \({\mathrm {reg}}_{M}^{{\mathcal {P}}^+|^{\chi }(E^c \cap F)}(g) = 1\). Using SEP and the choice of \(\delta \), we must have
Clearly,
Thus,
violating the second condition of SEP. Therefore, by Theorem 4.3, Axiom 1 does not hold.
Finally, suppose that condition (c) in rectangularity does not hold. Then for some non-negative real vector \(\theta \in {\mathbb {R}}^{|S|}\),
We construct a decision problem D based on \((S,\Sigma )\). D has two stages: in the first stage, nature chooses a state \(s \in S\). In the second stage, the DM chooses an action from the set \(M = \{f,g\}\), with utilities defined as follows:
So we have
This means that SEP, and hence Axiom 1, is violated, a contradiction. \(\square \)
Rights and permissions
About this article
Cite this article
Halpern, J.Y., Leung, S. Minimizing regret in dynamic decision problems. Theory Decis 81, 123–151 (2016). https://doi.org/10.1007/s11238-015-9526-8
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11238-015-9526-8