Causality occupies a central position in human cognition . The AI community has been working for a long time on representations of causality. Informal descriptions of real-world phenomena in the form A causes B, are very common. B can be predicted or explained using A. In particular, causal modeling, whether applied in the context of economic systems or qualitative physics, has been the subject of a famous debate [15,16]. Causal modeling enables a complicated process to be decomposed into elementary sub-models and is thus very suitable for complex system analysis. Causal models provide expla- nations of the behavior of the modeled system that are close to human reasoning, which is completely excluded by the purely numerical calculus that constitutes the basis of control theory. Causal descriptions are the source of various reasoning modes useful for supervision: understanding, predicting, diagnosing and action advice . Causality plays an essential role in humandecisionmaking by providing a basis for choosing that action that is likely to lead to a desired result. Diagnosis, an important supervision aspect, is also typically a causal process because it consists in designating the faulty components that have caused, and can explain, the observed malfunctions. In this respect, the objective of the use of a causal model is to deal with the combinatorial explosion that arises with model-based diagnosis approaches [18–21].
At the very height of the international credit crisis and during the near collapse of the Icelandic banking system end of 2008, beginning 2009, the Icelandic politician Johanna Sigurdardottir attracted voters by promising to “end the era of testosterone”. The business man of the year in 2008 in Iceland was a woman, and after Sigurdardottir became prime minister in February 2009, half of her new ministers were chosen women. Furthermore the CEOs in two of the three largest banks were replaced by women. If anything this Icelandic tale shows that both from the public but also political side there is the conception that risky behaviour is an attribute deeply rooted in male decisionmaking, especially during crisis, and could be avoided by including more women into the decision process. Politically as well as economically it is therefore important to have some solid research telling us whether there is some truth in such a viewpoint. The specific purposes of this paper is indeed to introduce tools to understand the origins behind excessive risk taking formed through synchronization in collective decisionmaking, taking into account the gender issue.
accumbens, the pleasure center of the brain, which plays a role in sexual arousal and the heavy derived from certain recreational drugs. It is involved in reward, placebo effect, pleasure, laugher, addiction, fear and aggression. The emotions triggered by stimuli acting on the limbic system are not under the control of the cortex. The limbic system is tightly connected to the prefrontal cortex including the orbitofrontal cortex that is required for decisionmaking. The limbic brain has some autonomy from the cortex, in blocking any responsiveness of cortical areas, and in anaesthetizing the unpleasant feelings that do not reach the cortex but may stimulate certain areas of the cortex. This is important to note that the communication is thus unidirectional from the limbic system to the cortex, but the logical right hemisphere of the cortex may block the communication from the limbic system.
While evidence accumulation was present for both LV and HV stimuli, there were striking behavioral differences in terms of how evidence was accumulated between conditions. First, the speed of decision-making was modulated by the amount of accumulated evidence for HV (p,0.001) but not for LV information (p = 0.31), leading to a significant interaction (p = 0.025, Figure 1E). Second, the ‘‘impact’’ of each successive arrow on the final decision varied as a function of time and accumulated evidence only for HV trials. We defined the impact of an arrow as the extent to which the arrow changed the response proportion in the direction of the arrow (see Materials and Methods for details on the exact quantification procedure). We observed a monotonically increas- ing impact of arrows on the decision as a function of time for HV stimuli (p,0.001), while this modulation of time was only marginal for LV stimuli (p = 0.07), leading to a significant interaction (p,0.001, Figure 1F). Moreover, for HV arrows, the influence of the last arrow, defined as the extent to which it changed the response probability in the direction of the arrow, decreased linearly with the amount of previously accumulated evidence: the larger the amount of accumulated evidence, the less influence the last arrow had on the decision (p,0.001), as expected from a rational strategy of progressively disregarding the arrows once sufficient evidence is obtained (Figure 1B). This modulation of accumulation by prior evidence was absent for LV stimuli (p = 0.44), leading to a significant interaction (p,0.001, Figure 1G). Together, these results show that strategic effects on decision- making strongly depend on the awareness level of the stimuli. Interestingly, these results were not simply due to the possibility that, during HV trials, participants stopped performing the task after having observed a sufficient amount of arrows. A ‘‘logical counting’’ algorithm would not give any weight to the last arrow when two or four pieces of evidence had already been accumulated, since the last arrow cannot change the decision anymore. In our data, however, the last arrow did have a sizeable influence on the decision even when four pieces of evidence had already been accumulated (Figure 1G, red line, right data point). We further explored the relationship between decision-making performance and subjective confidence in a new group of 16 participants; this time we additionally asked them to rate their confidence of having responded correctly after every trial on a 6- point scale (1 = pure guess, 6 = 100% sure). Overall decision-
Here, we directly tested the causal role of DA in human explore/exploit behavior in a pharmacologi- cal, computational fMRI approach, using L-dopa (DA precursor) and haloperidol (DA antagonist) in a double-blind, placebo-controlled, counterbalanced, within-subjects design. Model comparison revealed that choice behavior was best accounted for by a novel extension of a Bayesian learning model ( Daw et al., 2006 ), that included separate terms for directed exploration and choice persev- eration. Modeling revealed that directed exploration was reduced under L-dopa compared to pla- cebo and haloperidol. In contrast, no drug effects were observed on parameters capturing random exploration (b) or perseveration (r). On the neural level, exploration was associated with higher activ- ity in the FPC, IPS, dACC, and insula, whereas exploitation showed higher activity in the vmPFC, OFC, PCC, precuneus, angular gyrus, and hippocampus, replicating previous studies ( Daw et al.,
For other sophisticated models or not necessarily of the aggregation type, like the GAI (Gen- eralized Additive Independence) model of (Fishburn, 1967), we are not aware of any general definition of an importance index. This paper precisely aims at filling this gap. Our approach is quantitative, as our aim is to define a numerical importance index, and therefore our starting point will be a numerical representation of the preference, rather than the preference itself. The strong point of our approach is that no specific assumption is needed on the numerical represen- tation. In particular, we can deal with continuous or discrete-valued attributes, although most of our effort will bear on the discrete case, the continuous case being obtained by interpolation. Also, we do not assume monotonicity, that is, the satisfaction of the decision maker does not necessarily increase nor decrease with the values of the attributes.
decision A ∪ B for BBA’s m 2 (·) and m 5 (·) can be interpreted as a no proper
decision, in the sense that A ∪ B is the whole universe of discourse, hence we are merely selecting anything (and discarding nothing). Such kind of no proper decision may however be very helpful in some fusion systems because it warns that input information is not rich enough, and that one needs more information to take a proper decision (by including more sensors or more experts reports in the system for instance). For symmetrical mass function, the decision drawn from the new proposed decision rule is consistent with what we can reasonably get because. To make a proper decision we will always need to introduce some possibly arbitrary additional constraints.
expectation in this range.
The fact that the seller cannot form a unique expectation completely alters the optimal strategy for the buyers: the buyers can now rationally bid anything in the interval from (1 + ρ 0 )S 0 to (1 + ρ)S 0 .
Recall that under the expected utility model, the seller would privately set a reservation price at the expected value of the asset, so that any buyer who submits a potentially winning bid would have a negative expected return. In the non-probabilistic setting, however, there is no unique expected value available to the seller, so buyers can justify making potentially winning bids. Moreover, since the seller can rationally set his reservation price to (1 + ρ 0 )S 0 (again in contrast with the expected
} A. Adla, P. Zaraté, J.L. Soubie. A Proposal of Toolkit for GDSS Facilitators. In : Group
Decision and Negotiation, Springer, Vol. 20 N. 1, p. 57-77, 2011
} A. Coulibaly, P. Zaraté, G. Camilleri. Implementing Voting Tools in GRUS (short
paper). In : International Conference on Decision Support Systems Technologies (ICDSST 2017), Namur, Blegium, 29/05/17-31/05/17, Isabelle Linden, Shaofeng Liu, Jason Papathanasiou, Christian Colot (Eds.), ICDSST 2017, p. 67-72, mai 2017.
DMUU is making a decision when there are many unknowns or uncertainties about the kinds of states of nature (a complete description of the external factors) that could occur in the future to alter the outcome of a decision. In other words, the consequence of the decision is highly affected by a host of conditions beyond one’s control, e.g., whether a farmer harvests his crop is highly dependent on weather conditions, or decisions about launching a new product could be influenced by market forces. Furthermore, based on the degree of uncertainty, DMUU has two subcategories: decisionmaking under strict uncertainty (DMUSU) and decisionmaking under risk (DMUR). “Strict uncertainty” means that the likelihood of various possible future conditions is quantitatively immeasurable. “Risk” assumes that DMs can assign a probability distribution to each state of nature based on their own experiences or historical frequencies. Five classic DMUSU methods are Laplace’s insufficient reason principle (Keynes, 1921), Wald’s Maximin (Wald, 1950), Savage’s Minimax regret (Savage, 1972), Hurwicz’s pessimism-optimism index criterion (Hurwicz, 1952) and Starr’s domain criterion (Starr, 1966). They were actively developed in the early 1950s. Each method proposes different ways of handling uncertainty. As the probability distribution of states of nature can be assigned in DMUR, the well-known DM methods of DMUR are the expected monetary value, the expected opportunity loss, the most probable states of nature and the expected utility (Taghavifard, Damghani, & Moghaddam, 2009).
Learning a sequential decision policy (Case 2, Naive aproach)
Problem: How to generalize this if the decisions shown in the
sample are random (i.e. not necessarily the optimal ones)?
Brute force approach: one could use open loop parts of sample
Keywords: Multiple criteria analysis, Multichoice game, Shapley value, Choquet integral.
The notion of importance of attributes or criteria in multicriteria decision aiding (MCDA) has always been central in the modelling and analysis of preferences. It is crucial for practitioners for many reasons. Once the model is learnt, the user needs to check the validity of the model, in particular, by checking whether the mean importance of criteria fits with his feeling. In the opposite case, he shall revise his preferences. If some criteria/attribute turn out to have a very small importance, they can be discarded in an approximate and simplified model. On the other hand, if the user wishes to have a synthetic explanation of the model, focusing on the few criteria/attributes having the largest mean importance provides a simple explanation of the model. Finally, the user is often asking for explanations of decisions made by the model. A possible approach is to compute the importance of each criterion for a particular decision (Labreuche and Fossier, 2018). It is instructive to compare this “local” importance with the mean importance. For instance, a mean importance higher than the local importance implies that the corresponding criterion has been weakened by other ones for that particular decision.
B. Decisionmaking in dilemma situations
As mentioned in the previous subsection, the ethical dilemma scenario is created moving VEH abruptly in just one time step (also, the x coordinate of P1 and P2 are changed to 20 and 22.5, respectively). At this point (where the AV is in the same state for all policies tested) the AV must deliberate about an action, using one of the deliberation processes pro- posed in subsection IV-B. The velocity difference between each road user (∆v) is not significant enough to be identified in the fatality probability versus ∆v graphic given by  and , thus the constant c vul used for the collisions will
CREST, ENSAE & Criteo AI Lab, Paris
February 15, 2021
We consider the scenario in which a company faces a sequence of decisions to make, such as whether or not to implement certain updates on their systems. Each such decision is usually made based on a so-called A/B test, estimating the potential benefits of the update (for simplicity, we focus on a single metric, such as the total benefits, the gross sales, etc.). A crucial feature is that the longer an A/B-test is run, the more accurate its prediction. On the other hand, shorter runs ensure that more updates can be tested out in the same amount of time.
Figure 3 – Illustration of MOORA method 4. Conclusion
Employing technologies and innovative systems aids development a body of knowledge. In area of decisionmaking theories and applications, evolution of decision support system allows experts and decision makers to get the chance of interfacing to a database and facilitating decision process. This short communication tries to present application of MCDM methods and their implementation in a decision support system. For this, a decisionmaking problem is composed and the solution then is validated and visualized by a system called STORMa. Decision problem is basically tends to assess robots for a specific usage with respect to some criteria. We have solved decision problem with utilization of MOORA and COPRAS techniques and tried to enhance the accuracy of the results with implementation of the problem by decision-making software. This study suggests the development of other MCDM tools like TOPSIS or WASPAS in order to get such improvement in MCDM class. In terms of group decisionmaking approach, the decisionmaking can be carried out by a team of decision makers to reflect different and integrated opinion of the participants. The proposed software and decision support system are enough flexible which enable a group of experts to decide efficiently. Therefore, it is possible to implement a group decisionmaking structure in order to bring an optimized perspective for future research direction.