GM與ANFIS機車跟車模式之比較

MOTORCYCLE-FOLLOWING MODELS OF GENERAL MOTORS (GM) AND ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM (ANFIS)

本文旨在探討具有跟車現象之機車行為特性，界定影響機車跟車行為之顯著因素，據以構建機車跟車模式。經實地錄影觀察發現，機車具有跟車現象的比率僅占所有觀察機車樣本之 13.8%。由機車與旁車間主動所呈現之相關資料分析與檢定結果發現，相對速率差、與前車間距、前車加速率係影響後（機）車跟車加速率之顯著因素。依據實地觀察資料，本文首先構建通用汽車 (GM)機車跟車模式，發現情況 1 （僅正前方有車）之迴歸判定係數 (R2) 僅為 0.20.0.29間，均方誤差根 (RMSE) 值為 0.73.0.81；情況 2（正前方及斜前方含左或右或左右同時有車）之值亦僅為 0.06.0.14間，RMSE值更達 0.89.0.97，均顯示 GM跟車模式配適度相當低，難以有效描述機車之跟車行為。因此本文另構建適應性類神經模糊推論系統 (ANFIS) 機車跟車模式，發現情況 1及情況 2之 RMSE值分別僅為 0.16及 0.34，其配適度比 GM跟車模式優越甚多；進一步以 Q-Q相關係數檢定發現，情況 1通過預測值與觀察值完全正相關之檢定，情況 2雖未通過完全正相關檢定，惟其統計量相當接近臨界值，顯示情況 2預測值與觀察值亦近乎完全正相關，說明 ANFIS跟車模式確能描述實際之機車跟車行為。

The main purposes of this paper are to investigate the characteristics of motorcycle flow in a mixed traffic, to identify significant factors affecting the motorcycle-following behaviors, and to construct models that can properly describe the relationship between motorcycle acceleration rates and these factors. A field observation is conducted and outcomes shows that only 13.8% of the overall samples reveal a motorcycle-following phenomenon. Statistical tests show that the significant factors affecting motorcycle-following behaviors include relative speed, space headway between a motorcycle and its leading vehicle, and acceleration rate of leading vehicle. General Motors (GM) five- generation models are firstly attempted to explain the motorcycle’s following behaviors in two cases: (1) only one leading vehicle in front; (2) two or more leading vehicles in front and neighboring-front (including either left-front, right-front, or both). The rather low values of the coefficient of multiple regression determination (R2=0.20~0.29 and 0.06~0.14 for both cases) and relative large root-mean-square-error values (RMSE =0.73~0.81 and 0.89~0.97 for both cases) imply that all of the GM models have poorly described the motorcycle-following behaviors. Therefore, further propose a motorcycle-following model by incorporating the adapwe tive neuro-fuzzy inference system (ANFIS) with those significant factors that affect the motorcycle-following behaviors. Compared with the GM models, the ANFIS model outperforms with much smaller RMSE values (0.16 and 0.34 for both cases). Moreover, the Q-Q plot correlation coefficient tests also reveal that the predicted acceleration rates have a highly strong positive correlation with the observed acceleration rates in case (1) and a strong positive correlation in case (2). It suggests that our proposed ANFIS model can satisfactorily capture the nature of motorcycle-following behaviors in a mixed traffic.