This paper proposes a multi-objective ensemble-discrimination reinforcement learning method using mixed-parameterized actions to address the multi-objective compatibility problem in autonomous driving. Existing reinforcement learning methods struggle to achieve multi-objective compatibility in complex driving scenarios due to their single-valuation network and single-type action space structure. The proposed method addresses these challenges by utilizing an ensemble-discrimination method that focuses on different objectives through independent reward functions. Furthermore, by incorporating mixed-parameterized action space structures, it generates driving behaviors that encompass both abstract guidance and concrete control commands. Finally, it develops an uncertainty-based search mechanism that supports mixed actions to accelerate the learning of multi-objective-compatible policies. Experimental results in multi-lane highway scenarios, both simulator-based and on the HighD dataset, demonstrate that the proposed method efficiently learns multi-objective-compatible autonomous driving in terms of efficiency, behavioral consistency, and safety.
