Deep learning methods for sequential decision-making in dynamic systems

This doctoral dissertation investigates the development, improvement, and evaluation of Deep Learning methods for sequential decision-making in complex and dynamic systems. Such systems are characterized by temporal dependencies, uncertainty, noisy observations, partial observability, non-stationary conditions, and continuously changing environments. The main objective of the dissertation is to design intelligent models that can extract informative representations from sequential data, adapt their decisions over time, and operate reliably under real-world constraints. The proposed methods combine Deep Learning, Reinforcement Learning, imitation learning, uncertainty estimation, and physics-informed learning, with applications in algorithmic trading, autonomous driving, robotic manipulation, and dynamic-system modeling. In algorithmic trading, the dissertation introduces a series of learning-based methods, including TraderNet-CR, UNSURE, VPM, and ViT², for processing financial time series and supporting cryptocurrency trading decisions. These approaches investigate the integration of market representation learning, risk-sensitive decision-making, uncertainty estimation, and transferable predictive models. Particular emphasis is placed on improving robustness under volatile market conditions and reducing the risks associated with unreliable or overconfident predictions. In autonomous driving, several architectures are developed for perception, driving-policy learning, and adaptive vehicle control. The proposed CIL++, CILRL++, and CILRLv3 methods combine Conditional Imitation Learning with Reinforcement Learning to improve driving performance, generalization, and adaptation to previously unseen situations. Additional approaches investigate multimodal perception and temporal modeling, while NADAS addresses adaptive driving and Advanced Driver-Assistance Systems under computational and real-time constraints. The dissertation also examines the control of a human-sized InMoov robotic hand. Deep Learning models are developed for translating visual information and natural-language commands into robotic actuator movements, enabling human-to-robot motion transfer and speech-driven manipulation. These methods demonstrate how heterogeneous sensory inputs can be transformed into coordinated robotic actions. Finally, the dissertation studies Physics-Informed Neural Networks for modeling dynamic systems governed by differential equations. Methods such as GPS, GENESIS, REPS, and GENERAL introduce adaptive and stochastic sampling strategies that focus training on informative regions of the physical domain. These approaches improve convergence and predictive accuracy, particularly in problems involving complex dynamics and steep solution gradients. Overall, the dissertation demonstrates that robust sequential decision-making requires models that combine data-driven representation learning with uncertainty awareness, adaptability, domain knowledge, and application-specific constraints. The results contribute toward the development of intelligent systems that are more accurate, efficient, scalable, and reliable when deployed in complex real-world environments.