A reinforcement learning model for personalized driving policies identification
|نوع نگارش مقاله||
scopus – master journals – JCR
۴٫۲۷۶ در سال ۲۰۲۰
۲۶ در سال ۲۰۲۱
۰٫۹۰۱ در سال ۲۰۲۰
|شاخص Quartile (چارک)||
Q1 در سال ۲۰۲۰
خرید محصول توسط کلیه کارت های شتاب امکان پذیر است و بلافاصله پس از خرید، لینک دانلود محصول در اختیار شما قرار خواهد گرفت و هر گونه فروش در سایت های دیگر قابل پیگیری خواهد بود.
فهرست مطالب مقاله:
Optimizing driving performance by addressing personalized aspects of driving behavior and without posing unrealistic restrictions on personal mobility may have far reaching implications to traffic safety, flow operations and the environment, as well as significant benefits for users. The present work addresses the problem of delivering personalized driv- ing policies based on Reinforcement Learning for enhancing existing Intelligent Transportation Systems (ITS) to the benefit of traffic management and road safety. The pro- posed framework is implemented on appropriate driving behavior metrics derived from smartphone sensors’ data streams. Aggressiveness, speeding and mobile usage are consid- ered to describe the driving profile per trip and are presented as inputs to the Q-learning algorithm. The implementation of the proposed methodological approach produces per- sonalized quantified driving policies to be exploited for self-improvement. Finally, this paper establishes validation measures of the quality and effectiveness of the produced poli- cies and methodological tools for comparing and classifying the examined drivers.
|بخشی از متن مقاله:|
Optimizing driving performance by addressing personalized aspects of driving behavior is a focal research area, which may have far reaching implications to traffic safety and operations, environment, as well as significant benefits for users. The above renders necessary the processing of large-scale data explicitly describing the driving task. To this end, smart- phones can constitute an alternative data collection tool for driving behavior analysis (Weiser et al., 2016). Apart from wide diffusion, smartphones demonstrate several additional advantages, making instrumenting a vehicle for data collection accessible to the general public, academia and the industry, compensating for the current lack of dedicated equipment, like on-board diagnostics (OBD) devices (Johnson and Trivedi, 2011; Briante et al., 2014). They can form a non-intrusive environ- ment for continuously collecting rich and more granular data on the actual driving task, due to a variety of plugged in sensors (Zhao, ۲۰۰۰; Wahlström et al., 2015; Vlahogianni and Barmpounakis, 2017). Moreover, their communication features enable the connection with cloud based services and platforms (Lane et al., 2010), while the information collected, when exchanged between vehicles or with the road infrastructure, can contribute to a wide range of Intelligent Transportation Systems (ITS) services (Briante et al., 2014). For example, smartphones have recently penetrated to usage-based motor insurance (UBI)
schemes. The concept of UBI systems entails the idea of adjusting insurance costs according to parameters influencing the probability of accident involvement, like total exposure (pay as you drive – PAYD) or individual driving behavior (pay how you drive – PHYD) (Litman, 2004; Tselentis et al., 2016). UBI schemes come to revolutionize traditional insurance approaches with the implementation of advanced custom-made feedback mechanisms to persuade users to understand their driving behavior and raise awareness on the effects of safe driving (Toledo et al., ۲۰۰۸; Lane et al., ۲۰۱۰; Birrell et al., ۲۰۱۴; Baecke and Bocca, ۲۰۱۷; Tselentis et al., 2017).
Although conceptually the use of smartphones for the above-mentioned applications is utterly justified, the complexities that come with the smartphone-based driving applications are numerous and methodologically difficult to be tackled. Apart from the difficulties in delivering meaningful data on the travel mode, driving patterns, harsh maneuverings and phone interaction while driving (Handel et al., ۲۰۱۴; Vlahogianni and Barmpounakis, 2017; Kanarachos et al., ۲۰۱۸), robust approaches should be developed to identify driving profiles. These profiles can lead to meaningful, personalized driving poli- cies useful for raising the driver’s awareness, or monitoring and ranking drivers with respect to their performance. Delivering driving profiles is a research field that is constantly gaining a lot of attention (Júnior et al., 2017; Yu et al., 2017; Mantouka et al., 2018). However, the manner these profiles may be combined to produce meaningful – for both the user and the system
– policies is still at a native stage.
The present work aims to deliver a methodology that automatically formulates personalized quantified driving policies, to be disseminated to drivers as recommendations for future behavior without posing unrealistic restrictions on personal mobility. More precisely, the recommendations produced in no way restrict drivers by redefining their destination, redirect- ing their pre-chosen route or even affect their travel times. They only lead the users to achieve more prudent driving with less redundant, risk prone and, in the cases of speeding and mobile usage, illegal maneuvers. The methodological framework is founded on Reinforcement Learning (RL). Driving behavior can be extremely diverse among different drivers and during different trips of the same driver. This volatility of driving metrics is attributed to various parameters. This research defines the concept of optimum driving and seeks for the driving patterns, or else the recurring driving maneuvers, that should be alleviated for behavioral improvement to be achieved. The degradation of the quality of trips is expressed through their divergence from optimum driving. The proposed methodology is implemented in a set of 189 drivers monitored for over a year in a naturalistic driving experiment.
The remainder of this paper is structured as follows. The next sections provide a description of data and analyze the prin- ciples of RL and the Q-learning algorithm. Following, the methodology is outlined and the findings of this research are cited, as well as the validation measures and tools for classifying the examined drivers. Finally, findings are discussed and conclu- sions are drawn, while suggestions for future research are provided.
لطفا پیش از ارسال نظر، خلاصه قوانین زیر را مطالعه کنید: فارسی بنویسید و از کیبورد فارسی استفاده کنید. بهتر است از فضای خالی (Space) بیشازحدِ معمول، شکلک یا ایموجی استفاده نکنید و از کشیدن حروف یا کلمات با صفحهکلید بپرهیزید. نظرات خود را براساس تجربه و استفادهی عملی و با دقت به نکات فنی ارسال کنید؛ بدون تعصب به محصول خاص، مزایا و معایب را بازگو کنید و بهتر است از ارسال نظرات چندکلمهای خودداری کنید.