Efficient Tactical Decision Making for Trucks in Highway Traffic with Deep Reinforcement Learning

Licentiate thesis, 2026

This thesis investigates tactical decision making for autonomous heavy-duty trucks in highway traffic using deep reinforcement learning, with a particular emphasis on optimizing safety, efficiency and costs. The key aspects of decision making include Adaptive Cruise Control (ACC) and lane changes, which strongly influence energy consumption, travel time, and traffic interactions. To support a systematic study of this problem, we develop a scalable traffic model on a simulation platform, providing a controlled and extensible environment for autonomous truck driving in multi-lane highways.

We propose a hierarchical control architecture in which reinforcement learning is used for high-level tactical decision making, while low-level tactical actions are handled by physics-based controllers. This separation is found to improve the performance by reducing safety risks and facilitates the integration of learning-based decision making with established control methods. A realistic reward function is designed to jointly capture safety, efficiency, and operational costs, and advanced training strategies such as curriculum learning are investigated to handle conflicting objectives within a scalarized framework.

We further explore a multi-objective reinforcement learning formulation to explicitly represent trade-offs between competing objectives, enabling the learning of interpretable Pareto frontiers. The results demonstrate that learning-based tactical decision making policies can achieve meaningful trade-offs between safety and various operational costs in abstracted highway scenarios, and that multi-objective formulations provide valuable insight into the structure of these trade-offs. Overall, this work contributes to methodological foundations and evaluation tools for economically meaningful and extensible learning-based tactical decision making for heavy-duty trucks.