Runxuan Liu, Xianhao Ou, Xinyan Ma, Jiyuan Wang, Jiafeng Liang, Jiaqi Li, Tao He, Zheng Chu, Rongchuan Mu, Zekun Wang, Baoxin Wang, Dayong Wu, Ming Liu, Shijin Wang, Guoping Hu, Bing Qin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Multiple response optimization (MRO) aims to obtain an optimal solution by optimizing several responses simultaneously for quality improvement. The preference parameters in the traditional MRO are exact real numbers. However, qualified uncertainties of marginally qualified responses can lead to preference parameters, such as the lower and upper bounds, to obtain interval forms. To solve the MRO involving conflicting responses using desirability functions under interval preference parameter, an intuitionistic fuzzy desirability function method (IFDFM) is proposed, which includes the experiments and data regression phase, the intuitionistic fuzzy desirability function set (IFDFS) construction phase, the optimization and decision-making phase. Compared with the existing desirability function methods, in the proposed method, the IFDFSs are introduced. Namely, after evaluating interval preference parameters with marginally qualified responses, the IFDFSs are constructed to measure the qualified uncertainties of multiple response performances based on the membership, non-membership and hesitant degrees. The proposed IFDFM is verified on the tire-tread compound problem. The results show that the IFDSs of the optimal solution obtained by the IFDFM can better measure the desirability uncertainties of the marginally qualified responses. By comparing results of the IFDFM under different combinations of marginally qualified responses, the influence orders of responses on optimal solutions are obtained, providing important information for developing quality improvement strategies.

This study considers a response surface methodology (RSM) variation in which a response has multiple central tendencies (MCTs) that can have multiple influences on a sheet metal part. This is formulated as a set response surface (SRS) problem, which, in industrial practice, is studied using the thinning ratio example. The set RSM (SRSM), consisting of three phases, is proposed to solve the SRS problem. The first phase is the problem definition and regional division phase, where based on the analysis of MCTs and response influence, a sheet metal part that needs quality improvement is divided intoqregions. The second phase is the experiment and data regression phase. The third phase is the optimization and interactive decision phase, where the condition relaxation strategy (CRS) is proposed, and relaxation is obtained based on the barrier and engineering requirement analysis. The optimization models are constructed for the (q+1)-level optimal solutions using the CRS. The proposed SRSM is verified by tests on the wrinkle and crack problem of the inner plate of the back door. A possible optimization model combination and trend analysis strategy is proposed to solve the CRS challenge for higher dimensions in the tendencies.

Human action recognition and prediction with skeleton-based data has been widely applied in intelligent robots and machine vision. Current advanced methods mainly focus on using recurrent neural network (RNN) to predict human motion based only on skeleton data. However, objects that interact with the human body will have a great impact on human behaviors. Extracting body information without considering environmental objects constraints reduces the accuracy of behavior prediction. In this paper, we propose a scene-perception architecture, which implements an end-to-end human behavior prediction task based on both objects and human skeleton. We propose a scene-perception Graph Convolutional Network (SPGCN) to formula the natural constraint between skeleton nodes and objects in the scene to predict human postures. SPGCN is composed of graph convolutional network with scene-aware and RNN, GCN is used to learn the dependence relationships between human and objects. Then, the learned features are fed into RNN to predict future pose and action labels. We evaluate SPGCN on CAD-120 dataset. Experiments show that our proposed method achieves promising results compared with the state-of-the-art methods.