Meta-Cognitive Digital Twin Agents for Autonomous Vehicles: Self-Simulating Intelligence and Meta-Agent Society Framework

Abstract

Autonomous vehicles (AVs) in Internet-of-Things (IoT) ecosystems require adaptive, self-reorganising intelligence to handle dynamic driving behaviours, sensor variability, and safety threats. This paper proposes the Meta-Cognitive Digital Twin Agent (MCDTA) framework, integrating a society of specialised agents with a reliability-driven meta-agent and a cognitive digital twin layer. Six heterogeneous base agents (Random Forest, Extra Trees, Gradient Boosting, SVM, Logistic Regression, AdaBoost) are autonomously monitored, ranked by a composite Reliability Score, and reorganised by a meta-agent that removes weak classifiers and spawns specialists when collective strength falls below threshold. A cognitive digital twin then simulates four candidate actions, selecting the safest under predicted future risk. On the Kaggle Driving Behaviour dataset (n=5,620), MCDTA attains Accuracy=0.9821, F1=0.9819 and ROC-AUC=0.9981, exceeding seven baselines by 1.70–8.90 points. The framework advances autonomous-systems intelligence through self-organising agentic AI, meta-cognitive reliability assessment, and pre-execution risk simulation.