arXiv:2409.04462

New parametric identification method for a preference model

J Renaud, M Camargo, L Morel, C Fonteix

제출 2024-09-02

원문 초록

This article presents a contribution to multi-criteria decision support intended for industrial decision-makers in order to determine the best compromise between design criteria when working on risky or innovative products. In (RENAUD et al. 2008) we used the OWA operator (Ordered Weighted Average), a well-known multi-criteria analysis technique introduced by (YAGER 1988). The interest of this aggregation method is, beyond its ease of use, its ability to evaluate a product according to a single scale. When using the OWA method, the choice of criterion weights and their values remains an important and delicate operation. Indeed, the weights are not fixed by criterion but according to a level of utility (FISHBURN 1967). The weights can, in addition, be determined using different methods. A traditional approach consists in estimating the weights from an a priori selection of the most representative samples by an expert. We propose a new approach based on applied D-Optimality to determine the best sample. The results of the two approaches are compared. Indeed, in some multi-criteria decision problems, it is difficult to choose the method that best describes the behavior of the decision maker. As said above, the first part of this paper presents an original method based on D-optimality to determine the most representative set of samples to determine the decision parameters. This method has been applied and validated in an OWA approach. It has been shown that the accuracy and reliability of the method have been improved. Since the OWA application has been improved, the other goal of this paper is to verify whether the real decision maker has an OWA-based behavior or not. To achieve this, the D-optimality approach is applied to the MAUT (Multi-Attribute Utility Theory) approach to find the best set of samples. The results of the two methods are compared. Both approaches show that, while OWA better simulates the decision maker's behavior, there is still a gap between its scores and that of OWA. The question here is how to improve the accuracy of the score estimated by OWA compared to those given by the expert. Subsequently, a hybrid approach is tested, such as a linear combination of the MAUT and OWA approaches. The results obtained with this combination were found to be more accurate than those of OWA. However, we also tested the CHOQUET discrete integral approach. In other words, it is possible to find a model capable of better describing the decision maker's behavior.

함께 살펴볼 논문

공통 태그를 근거로 연결했습니다. 실제 인용 관계나 논문의 우수성을 뜻하지 않습니다.

arXiv:2506.23921v5cs.CL공통 태그 Attribute공식 코드 · Papers with Code

대규모 언어 모델에서 진실의 트릴레마

The Trilemma of Truth in Large Language Models

일반 대중은 대규모 언어 모델(LLM)에 인간과 유사한 특성을 부여하며, LLM이 특정 정보를 '알고' 있다고 가정하는 경우가 많다. 실제로 LLM은 학습 중 보존된 정보를 내부적 확률적 지식으로 인코딩한다. 본 연구는 해당 지식의 진위를 탐칭(probing)하는 기존 방법들을 검토하고, 세 가지 결함 있는 근본 가정을 밝힌다. 이러한 결함을 해결하기 위해, 다중 인스턴스 학습(MIL)과 conformal prediction을 결합한 다중 클래스 탐칭 프레임워크인 sAwMIL(Sparse-Aware Multiple-Instance Learning)을 제안한다. sAwMIL은 LLM의 내부 표현(representation)을 활용하여 진술을 참, 거짓, 또는 둘 다 아님으로 분류한다. 우리는 새롭게 구축한 세 개의 데이터셋을 사용하여, 기본형 및 채팅 기반 변형을 포함한 16개의 오픈소스 LLM에 대해 sAwMIL을 평가한다. 결과는 (1) 일반적인 탐칭 방법은 신뢰할 수 있고 전이 가능한(transferable) 진위 방향을 제공하지 못하며, 일부 설정에서는 제로샷(zero-shot) 프롬프팅(prompting)보다 성능이 낮음을, (2) 진실과 거짓은 대칭적으로 인코딩되지 않음을, (3) LLM이 참과 거짓 모두와 구별되는 제3의 유형의 신호를 인코딩함을 보여준다.

토론

이 논문과 연결된 포럼 글이 아직 없습니다.