Enhancing Discretionary Lane Change Decision Models Based On Fuzzy Inference Systems

A Lane Changing Decision Model (LCDM) mimics a human driver’s {yes, no} decision to start to move the vehicle from the existing lane to the target lane. LCDM is a critical component of semi- and fully automated driving systems. Recent research has found that the Fuzzy Inference System (FIS) is a promising approach to implementing LCDMs. This dissertation reviewed and identified several challenges in the development of existing FIS models to make the {yes,no} decisions for Discretionary Lane Changes (DLCs). This dissertation then addressed these challenges by developing a refined FIS framework through two key improvements. The first improvement fine-tuned fuzzy rules and applied advanced composition methods and defuzzification techniques. The second improvement included relative speeds as additional input variables. The inclusion of the relatively new highD data set, in addition to the established NGSIM data set, tested the FIS framework in congested and uncongested conditions. Each data set was re-processed to have approximately equal number of {yes, no} data points. The test results showed that the refined FIS framework, using only gap and distance variables as inputs, produced decisions that matched the observed maneuvers reasonably well. Separating the fuzzy inference rules into distinct {yes} and {no} groups and performing compositions of the rule groups separately further improved the FIS framework’s decision performance. Despite these refinements in the FIS framework, the Gene Expression Programming (GEP) model still produced a higher match in the {yes, no} decisions with the NGSIM test data set. When relative speed variables were incorporated alongside the gaps and distance inputs, the refined FIS framework demonstrated improved performance. The evaluations using the NGSIM (congested) and highD (uncongested) datasets confirmed that the enhanced FIS framework with relative speed variables as additional inputs consistently outperformed the gap- and distance-based version. The improvement was particularly evident in the highD data set, underscoring the importance of relative speeds for DLC decisions in uncongested conditions. In the tests with the NGSIM data set, the enhanced FIS framework outperformed the Gap Acceptance Model (GAM) and the GEP model. A sensitivity analysis on the levels of fuzzy sets of the relative speed variables showed that, while a reduction in the fuzzy set levels (number of linguistic variables) from six to four had little effect on the FIS’s decisions, a further reduction to fuzzy sets of two led to a noticeable fall in performance. These findings have established an enhanced FIS framework for lane changing decision modeling, with strong potential for integration into traffic simulation and autonomous vehicle systems.