RUM: Random Regret Models
Project details
Duration
09.2026-08.2017
Sponsor
external page SERI Swiss Government Excellence Postdoctoral Scholarship
Staff
Prof. Dr. Kay W. Axhausen and F. Belgiawan
Summary
When facing several alternatives, people are often assumed to choose the alternative which maximizes their utilities. This concept is widely known as random utility maximization (RUM). In transportation research, one of the most popular models, the multinomial logit (MNL) model, is based on this idea, e.g. for modeling mode choice. The choice in this model does not consider the relative position of the alternatives. It is context-free.
Recently there is a growing interest in an alternative modeling approach, random regret minimization (RRM), which considers the relative performance of the alternatives and is therefore context-dependent. In RRM, an individual, when choosing between alternatives, is assumed to minimize anticipated regret as opposed to maximize his/her utility. There are three variants of RRM, the classical CRRM, the µRRM, and the P-RRM. There is also a further approach called relative advantage maximization (RAM) turning the idea around and focusing on the gains. We finally test the mirror image of the RAM approach and find interesting differences.
We compare MNL with the four mentioned alternatives. We use stated choice data sets collected by the Institute for Transport Planning and Systems, ETH Zurich, which include mode choice, location choice, parking choice, carpooling, car-sharing and other experiments. We compare the performance of those five models by their model fit (Final LL, hit rate, and prediction accuracy). We also present a comparison of their resulting values of travel time savings (VTTS), plus travel time and cost elasticities.
Looking at the model fit, we found that RAM outperforms the other models in five cases, whereas the PRRM does so in two cases and µRRM only for one case. Only comparing the MNL and CRRM, we found that only in two cases MNL outperforms CRRM. Regarding the prediction rate, we found an interesting result: in more than 80% of the cases almost all models predict the same outcomes. This indicates that whatever model we use, we might end up obtaining a similar predictions. Still, the VTTS and elasticities vary substantiallym which is relevant for cost-benefit analysis or simplified modelling approaches.