Discrete choice model (DCM) is a classical framework for modelling an individual’s travel choice. However, its oversimplified architecture of utility function may limit its performance when faced with a complex decision process. In this paper, we develop a new framework called multi-task neural network and additive gaussian process based discrete choice model (MNNAGP-DCM). Specially, the multi-task neural network (MNN) is used to learn the representation of individual characteristics, while the additive Gaussian process regression (AGP) process is utilized to enhance flexibility of utility function. In multi-task neural network, the sub-learners learn the taste parameters between individuals’ characteristic in each alternative, while the global bias term is used to learn the cross effect between alternatives. The additive GPR framework is employed to substitute the linear term in utility function with a nonparametric probability framework. Additive GPR enables the modelling of nonlinearity, threshold effects and uncertainty, thereby providing a more comprehensive perspective on the decision-making process. Moreover, when combined with DCM, the GPRs become intractable. To address this, we employ variational inference to construct a tractable lower bound, thereby transforming the original model into a tractable one. Then MNNAGP-DCM can be optimized by gradient based algorithms such as Adam. The proposed model is tested on the open-source dataset and benchmarked with standard MNL, Mix-logit, XGBoost, TasteNet-MNL, MNN-DCM and MNNSGP-DCM. Results show that MNNAGP-DCM can not only capture individuals’ heterogeneity but also can learn the nonlinearity in utility function, showing great superiority in terms of predictability. Our model can also provide interpretable result with taste parameters and the fitted GPR models, while quantifying uncertainty through GPR’s probability framework.
扫码关注我们
求助内容:
应助结果提醒方式:
