Neuro-Memory Fuzzy Inference System for Mimicking Human-like Car Following Behavior

• The manuscript is 20 pages long.Source: Md. Hadiuzzaman · p. 1
• The submission date is August 1, 2026.Source: Md. Hadiuzzaman · p. 1
• The metadata identifies a corresponding author but does not name them.Source: Md. Hadiuzzaman · p. 1

• Car-following models represent vehicle interactions through variables such as relative speed, spacing, and velocity, and support applications including ADAS and ADS.These models are used to represent driver behavior and transfer human skills to intelligent systems.Source: INTRODUCTION · p. 3
• Earlier psycho-physical and machine-learning approaches expanded car-following models, but the psychological processes underlying driver behavior remain difficult to capture comprehensively.The literature describes perceptual thresholds, risk, distraction, and neural and deep-learning models as prior approaches.Source: INTRODUCTION · p. 3
• Human-like CAV decision-making remains a motivation because human drivers consider empathy, social context, and situational awareness in unpredictable scenarios.The paper notes that technically safe decisions may still affect stress among human drivers or pedestrians.Source: INTRODUCTION · p. 3
• NeMeFIS asymmetrically models acceleration and deceleration by integrating procedural, working, episodic, semantic, and declarative memory.It links memory functions to external driving variables through metaheuristic optimization and assesses them with factor analysis and p-value analyses.Source: INTRODUCTION · p. 3
• The paper proceeds from literature review and architecture to data and analysis, then model evaluation and conclusions.Source: INTRODUCTION · p. 3

Psychology-uninformed Car-Following Models

• Traditional stimulus-response models and later stability-focused models largely lacked psychological dimensions or behavioral depth.The review identifies Prospect Theory as an exception that modeled risk-taking under ambiguity.Source: Psychology-uninformed Car-Following… · p. 4
• Machine-learning approaches captured nonlinearities and latent psychological factors, but reported limits included validation, explicit reaction-time modeling, data demands, and fuzzy-rule complexity.The review describes neural networks, RBFN and fuzzy ARTMAP, LSTM, and LOLIMOT approaches.Source: Psychology-uninformed Car-Following… · p. 4

Psychology-informed Car-Following Models

• Psychology-informed models introduced memory functions and perceptual thresholds, while the reviewed Wiedemann model lacked cognitive components and could switch perceptual regimes without stimuli.Source: Psychology-informed Car-Following Models · p. 4
• Other approaches incorporated fuzzy inference, perception and vehicle-dynamics randomness, memory-informed behavior, or decision-making under uncertainty.The review describes contributions by Chakroborty and Kikuchi, Fadhloun and Rakha, Huang, Hamdar, and Li.Source: Psychology-informed Car-Following Models · p. 4
• Explicit cognitive models addressed decision regimes, driving styles, perception-memory-decision structure, fatigue, following defense, and anticipation delays.The review notes that the perception-memory-decision model lacked empirical validation.Source: Psychology-informed Car-Following Models · p. 4
• The review identifies interpretability, data demands, and realistic psychological integration as remaining limitations.Source: Psychology-informed Car-Following Models · p. 4

• Hierarchical fuzzy systems address the curse of dimensionality by decomposing high-dimensional problems into interconnected, low-dimensional fuzzy units.Outputs from one level become inputs to later levels, supporting scalability and reducing complexity.Source: Hierarchical Fuzzy Systems · p. 4
• The architecture begins with real input variables and fuzzy variables, with m denoting the total number of fuzzy rules.Source: Hierarchical Fuzzy Systems · p. 4
• At successive levels, a real input and the previous fuzzy-unit output are combined, and outputs are calculated recursively.Source: Hierarchical Fuzzy Systems · p. 5
• Input and output membership functions may use different forms, including bell-shaped and Gaussian functions.Source: Hierarchical Fuzzy Systems · p. 5
• Membership-function parameters a_i,j, b_i,j, and c_i,j correspond to input X_i, while Q_i,j denotes normalized firing strength.Source: Hierarchical Fuzzy Systems · p. 5

The Magnitude Model: Neuro-Memory Fuzzy Inference System (NeMeFIS)

• NeMeFIS links Procedural, Working, Episodic, and Semantic FIS subsystems to Declarative memory, then combines Declarative and Procedural outputs in Combined Memory.The architecture also includes a Combined Memory feature vector and subsystem output for the final decision representation.Source: The Magnitude Model: Neuro-Memory… · p. 5
• During training, tuned fuzzy rules and membership functions encode habitual, short-term, event-specific, generalized, and integrated driving responses.At inference, current inputs activate rules in proportion to their match with stored prototypes.Source: The Magnitude Model: Neuro-Memory… · pp. 5–6
• The study collected data on Arterial, Collector, and Rural Highway corridors in Dhaka, Bangladesh.The corridors primarily serve private passenger cars, rickshaws, and inter-district buses, respectively.Source: The Magnitude Model: Neuro-Memory… · p. 6

Experimental Set-up and Flight Plan

• Drone video was captured at 30 FPS during 9:00–10:00 AM on one weekday at each of the three study corridors.The specified dates were January 8, 2019 (AR), January 30, 2023 (CR), and February 18, 2023 (RH).Source: Experimental Set-up and Flight Plan · p. 6
• Data collection avoided extreme fog, used drone altitudes of 23–27 meters, and included camera calibration and skew-axis correction.The viewing angle was approximately 17–18 degrees, with flights launched at least 20 meters from the road.Source: Experimental Set-up and Flight Plan · p. 6
• Road-user detections provide centroid-based positions that are tracked over time to form trajectories for vehicle interactions.Measurements are made between subject and lead vehicles; the passage describes a trajectory as a zipped collection of points over the desired interval.Source: Experimental Set-up and Flight Plan · pp. 6–7
• Equations (7)–(15) measure speed, acceleration or deceleration, time headway, relative spacing and speed, lateral gap, time to collision, and risk perception.The text defines the inter-position time interval and Euclidean distance function used in the measurements.Source: Experimental Set-up and Flight Plan · p. 7

Data Analysis

• Speed correlates positively with relative speed during acceleration (0.103 in AR, 0.134 in CR) and nearly perfectly negatively during deceleration (–0.987 in AR, –0.988 in CR, –0.967 in RH).Relative speed has strong positive deceleration correlations: 0.826 in AR, 0.758 in CR, and 0.764 in RH.Source: Data Analysis · p. 7
• Spacing variables correlate positively with relative speed during acceleration and moderately negatively during deceleration, while angular variables have weak correlations in most cases.Risk perception is also weakly correlated across scenarios, suggesting a nuanced or indirect relation to moment-to-moment speed regulation.Source: Data Analysis · p. 7
• The researchers examined hysteresis patterns across trajectory pairs to ensure they conform to true following behavior.Hysteresis is used to identify trajectory pairs in following behavior [51].Source: Data Analysis · pp. 7–8

Integrating External Variables into Human Memory-Based Architecture

• Factor loadings support procedural spacing variables in AR (0.565 and 0.948), episodic spacing-change variables in RH (0.697 and 0.736), and semantic speed in CR (0.992).The study interprets these groupings as habitual monitoring, recalling recent spacing variations, and knowledge-driven anticipatory decisions, respectively.Source: Integrating External Variables into… · p. 8
• Procedural Memory represents habitual and intuitive safe-spacing maintenance, while Working Memory represents real-time spatial awareness and decision-making.The procedural grouping is associated with low uniqueness values; the working grouping has moderate factor loadings and active-task specificity.Source: Integrating External Variables into… · p. 8
• Episodic Memory represents interpretation of dynamic spacing changes through event-based memory, while Semantic Memory reflects generalized knowledge of speed and risk.The semantic description is associated with abstract cognitive processes and high factor loadings.Source: Integrating External Variables into… · p. 8
• Speed (Var 7) is significant in all 14 CR and RH cases, while spacing and spacing-change variables also have many p-values < 0.05.Angle-related variables generally have higher average p-values and fewer significant cases.Source: Integrating External Variables into… · p. 8

Training and assessment

• Training uses Sugeno-type fuzzy systems optimized with Particle Swarm Optimization and validation-set Pattern Search, then evaluates RMSE on held-out test data.The study reports convergence in all three corridors; 54 models were trained, tested, and validated using a 70%/15%/15% train/validation/test split.Source: MODEL DEVELOPMENT AND ASSESSMENTS · p. 9

Performance evaluation

• NeMeFIS performs best mainly in deceleration across corridors and vehicle types, with lower RMSE than LR, ANFIS, and LSTM in 81 of 162 comparisons.Its advantage is strongest for passenger cars, buses, motorcycles, and mixed traffic, particularly on collector roads and rural highways; acceleration gains on those corridors are smaller.Source: MODEL DEVELOPMENT AND ASSESSMENTS · pp. 9, 12, Table 4 · p. 10

Rule analysis

• Deceleration has greater fuzzy-rule complexity, especially for declarative and combined memory, while acceleration uses fewer rules under procedural memory.Declarative memory consistently has the highest rule counts, reflecting its role in complex cognitive reasoning.Source: MODEL DEVELOPMENT AND ASSESSMENTS · p. 12

Variable sensitivity

• Risk perception remains a dominant influence on acceleration sensitivity across scenarios, while spacing strongly affects acceleration at low normalized values.In arterial and rural settings, speed and relative speed also show noticeable sensitivity.Source: MODEL DEVELOPMENT AND ASSESSMENTS · p. 12, Figure 2 · p. 13

Variable importance

• Variable importance shifts by phase and corridor: risk perception leads acceleration in arterial and collector roads, while spacing cues dominate deceleration.For rural-highway acceleration, angular spacing is more critical for buses; during deceleration, lead spacing ranks highest on arterial roads, angular distance on collector roads, and relative spacing on rural highways.Source: MODEL DEVELOPMENT AND ASSESSMENTS · p. 12, Figure 3 · p. 15

Memory contributions

• Procedural memory dominates acceleration on arterial and collector roads, while semantic and declarative memory lead deceleration, with contributions varying by corridor.Semantic memory governs rural-highway acceleration; declarative memory guides deceleration there and on collector roads.Source: MODEL DEVELOPMENT AND ASSESSMENTS · p. 15, Figure 4 · p. 16

Numerical simulation

• Simulations across the three corridor types support model stability under controlled acceleration, deceleration, and shockwave-inducing conditions.Trajectories showed no FIFO violations; abrupt speed changes had minor oscillations that damped over time, while emergency-braking spikes arose from vehicle-entry boundary conditions.Source: MODEL DEVELOPMENT AND ASSESSMENTS · pp. 15–16, Figure 5 · p. 17

Transferability

• NeMeFIS was transferable to both US-101 and Argoverse 2 trajectory datasets, with considerable accuracy and sensitivity in each.The conclusion also describes the model as robust and transferable across UAV-based heterogeneous traffic data, homogeneous US-101 data, and Argoverse 2 CAV trajectories.Source: MODEL DEVELOPMENT AND ASSESSMENTS · pp. 16–17, Figure 6 · p. 17

FUNDING

• The Committee for Advanced Studies and Research at BUET supported the work through Grant No. 371(16).Source: FUNDING · p. 18