A Generative Neuro-Cognitive Architecture Using Quantum Algorithms for the Autonomous Behavior of a Smart Agent in a Simulation Environment

This study aims to develop a quantum computing-based neurocognitive architecture that allows an agent to perform autonomous behaviors. Therefore, we present a brain-inspired cognitive architecture for autonomous agents that integrates a prefrontal cortex–inspired model with modern deep learning (a transformer-based reinforcement learning module) and quantum algorithms. In particular, our framework incorporates quantum computational routines (Deutsch–Jozsa, Bernstein–Vazirani, and Grover’s search) to enhance decision-making efficiency. As a novelty of this research, this comprehensive computational structure is empowered by quantum computing operations so that superiority in speed and robustness of learning compared to classical methods can be demonstrated. Another main contribution is that the proposed architecture offers some features, such as meta-cognition and situation awareness. The meta-cognition aspect is responsible for hierarchically learning sub-tasks, enabling the agent to achieve the master goal. The situation-awareness property identifies how spatial-temporal reasoning activities related to the world model of the agent can be extracted in a dynamic simulation environment with unstructured uncertainties by quantum computation-based machine learning algorithms with the explainable artificial intelligence paradigm. In this research, the Minecraft game-based simulation environment is utilized for the experimental evaluation of performance and verification tests within complex, multi-objective tasks related to the autonomous behaviors of a smart agent. By implementing several interaction scenarios, the results of the system performance and comparative superiority over alternative solutions are presented, and it is discussed how these autonomous behaviors and cognitive skills of a smart agent can be improved in further studies. Results show that the quantum-enhanced agent achieves 2× faster convergence to an 80% task success rate in exploration tasks and approximately 15% higher cumulative rewards compared to a classical deep RL baseline. These findings demonstrate the potential of quantum algorithms to significantly improve learning and performance in cognitive agent architectures. However, advantages are task-specific and less pronounced under high-uncertainty, reactive scenarios. Limitations of the simulation environment are acknowledged, and a structured future research roadmap is proposed involving high-fidelity simulation validation, hardware-in-the-loop robotic testing, and integration of advanced hybrid quantum-classical architectures.