Search for: All records

Abstract In this study, we focus on estimating the heterogeneous treatment effect (HTE) for survival outcome. The outcome is subject to censoring and the number of covariates is high-dimensional. We utilize data from both the randomized controlled trial (RCT), considered as the gold standard, and real-world data (RWD), possibly affected by hidden confounding factors. To achieve a more efficient HTE estimate, such integrative analysis requires great insight into the data generation mechanism, particularly the accurate characterization of unmeasured confounding effects/bias. With this aim, we propose a penalized-regression-based integrative approach that allows for the simultaneous estimation of parameters, selection of variables, and identification of the existence of unmeasured confounding effects. The consistency, asymptotic normality, and efficiency gains are rigorously established for the proposed estimate. Finally, we apply the proposed method to estimate the HTE of lobar/sublobar resection on the survival of lung cancer patients. The RCT is a multicenter non-inferiority randomized phase 3 trial, and the RWD comes from a clinical oncology cancer registry in the United States. The analysis reveals that the unmeasured confounding exists and the integrative approach does enhance the efficiency for the HTE estimation. 

World model based reinforcement learning (RL) has emerged as a promising approach for autonomous driving, which learns a latent dynamics model and uses it to train a planning policy. To speed up the learning process, the pretrain-finetune paradigm is often used, where online RL is initialized by a pretrained model and a policy learned offline. However, naively performing such initialization in RL may result in dramatic performance degradation during the online interactions in the new task. To tackle this challenge, we first analyze the performance degradation and identify two primary root causes therein: the mismatch of the planning policy and the mismatch of the dynamics model, due to distribution shift. We further analyze the effects of these factors on performance degradation during finetuning, and our findings reveal that the choice of finetuning strategies plays a pivotal role in mitigating these effects. We then introduce AdaWM, an Adaptive World Model based planning method, featuring two key steps: (a) mismatch identification, which quantifies the mismatches and informs the finetuning strategy, and (b) alignment-driven finetuning, which selectively updates either the policy or the model as needed using efficient low-rank updates. Extensive experiments on the challenging CARLA driving tasks demonstrate that AdaWM significantly improves the finetuning process, resulting in more robust and efficient . 

World model based reinforcement learning (RL) has emerged as a promising approach for autonomous driving, which learns a latent dynamics model and uses it to train a planning policy. To speed up the learning process, the pretrain-finetune paradigm is often used, where online RL is initialized by a pretrained model and a policy learned offline. However, naively performing such initialization in RL may result in dramatic performance degradation during the online interactions in the new task. To tackle this challenge, we first analyze the performance degradation and identify two primary root causes therein: the mismatch of the planning policy and the mismatch of the dynamics model, due to distribution shift. We further analyze the effects of these factors on performance degradation during finetuning, and our findings reveal that the choice of finetuning strategies plays a pivotal role in mitigating these effects. We then introduce AdaWM, an Adaptive World Model based planning method, featuring two key steps: (a) mismatch identification, which quantifies the mismatches and informs the finetuning strategy, and (b) alignment-driven finetuning, which selectively updates either the policy or the model as needed using efficient low-rank updates. Extensive experiments on the challenging CARLA driving tasks demonstrate that AdaWM significantly improves the finetuning process, resulting in more robust and efficient . 

Many rapidly developing countries around the world are at a crossroads when it comes to transportation, air quality, and sustainability. Indeed, the challenges presented by vehicular growth in India have motivated the search for sustainable transportation solutions. One solution constitutes ridehailing services, which are expected to reduce car ownership and provide affordable means of transportation. Another key solution is the rise of electric vehicles (EVs), which are expected to reduce greenhouse gas emission and address the growing demand for sustainable urban mobility. Using a unique survey data set collected in 2018 from a sample of 43,000 respondents spread across 20 cities in India, this paper attempts to shed light on the factors that affect adoption of on-demand transportation services and EVs in India. In particular, not only does this paper consider the socio-economic and demographic variables that affect these behavioral choices, but the modeling framework adopted in this study places a special emphasis on representing the important role played by attitudes, values, and perceptions in determining adoption of on-demand transportation services and EVs. It is observed that attitudes and values significantly affect the use of on-demand transportation services and EV ownership, suggesting that information campaigns and free trials/demonstrations would help advance the adoption of sustainable transportation modes. The model results help in the identification of policy options and infrastructure investments that can advance a sustainable transportation future in India.