Abstract
Background. A generalization of the minimax regret criterion is represented as even the best-assurance minimax regret criterion comes inconsistent under instable evaluations of decision situations. Objective. The goal is to formulate the minimaximax regret criterion. Methods. Unlike the classic one, the generalized regret criterion is minimaximax which operates over generalized regrets. These regrets are found from a generalized decision function which is defined on a Cartesian product of a decisions’ set, a set of states, and a set of metastates. Metastate describes instability of the decision function whose values change through a set of metastates. The instability destroys assurance of minimaxed regrets found classically, so regrets are found over a generalized decision function. For this, utility evaluations are subtracted from the utility maximized across a decision set, or the loss/risk minimized across a decision set is subtracted from loss/risk evaluations. Then regrets are minimized under uncertainty across two dimensions of states and metastates, that is they are minimaximaxed. Results. The suggested minimaximax regret criterion allows finding an optimal decisions’ subset with not only regarding instability of the decision function, but also with reducing the initial decisions’ set more, unlike the ultimate pes-simism criterion without regrets (minimaximax/maximinimin). This especially concerns nonnegative utility matrices with many zeros. Conclusions. A ratio of a number of optimal decisions by the without-regret maximinimin/minimaximax to a number of optimal decisions by the minimaximax regret criterion decreased by 1 can be interpreted as a gain of applying the represented minimax regret criterion generalization. This gain fundamentally depends on whether sets of decisions, states, and metastates are finite or not. If they all are finite, then the gain depends on values in a three-dimensional regret matrix and its dimensions. It is surprising but the gain may be negative, that is finding regrets may come non-effective.
Highlights
Minimax regret criterion preferred to classic maximin/minimax is applied in a wide variety of decision making fields [1, 2]
An expected advantage of minimaximax regret criterion is that it will be always reducing the initial set of decisions X more
Finding regrets over a generalized decision function of three variables differs from classically finding regrets only with operating across that third dimension
Summary
Minimax regret criterion preferred to classic maximin/minimax is applied in a wide variety of decision making fields [1, 2]. Instability of the decision function destroys assurance of minimaxed regrets. Point evaluation is always much less reliable than interval evaluation (or, at least, multiple point evaluation), that implies multiple versions of the decision function. These aspects are equivalent to decision func*. A generalization of the minimax regret criterion is represented as even the best-assurance minimax regret criterion comes inconsistent under instable evaluations of decision situations
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
