Safe reinforcement learning in automotive

Thesis

This work applies Logically Constrained Reinforcement Learning (LCRL) frame-work for synthesizing policies for active monitoring and management of sensor modules in autonomous vehicles. LCRL allows synthesizing policies for unknown and continuous-state Markov Decision Processes (MDPs) such that a given linear time property is satisfied. We frame the problem of dynamic sensor module selection in automotive as an MDP and define the constraints as a Linear Temporal Logic (LTL) property. The LTL properties are used to guard and guide the MDP in finding an optimal policy. Defining a reward function over the state-action pairs of MDP based on the LTL property results in learning in a constrained environment. This approach leads to an improvement in performance and scalability for finding the right monitoring policies, as the safety properties enforces bounds on the search space.

Authors: Shah, A

• Alternative title: Safe automotive perception via deep logically-constrained reinforcement learning
• DOI: 10.5287/ora-bmbxx0nxx
• Type of award: MSc
• Level of award: Masters
• Awarding institution: University of Oxford
• Language: English
• Keywords: execution monitoring; adaptive active monitoring; safe RL; cautious RL; reinforcement learning (RL); autonomous driving; resource management with deep RL; deep RL
• Subjects: Artificial intelligence; Logic, Symbolic and mathematical; Deep learning (Machine learning); Reinforcement Learning; Automated vehicles