Supervisor: Radu Grosu
Safe learning algorithms for Autonomous Driving
Reinforcement Learning (RL) automatically synthesizes control policies, through the iterative improvement of an approximated solution, and the interaction of an agent with the operating environment. Despite the recent successes in many video games and control tasks, partial observability and its trial-and-error nature make it impossible to apply to real-world systems because of the safety implications. Many works have examined the delicate trade-off between safety and performance during training. However, partial observability characterizes most real-world systems and breaks the assumption of perfect knowledge that guarantees safe intervention. For this reason, this research proposal intends to investigate the design of safe-learning algorithms in the realm of partially observable realworld systems. We specifically target autonomous-driving applications as impactful testbed which offers high task-complexity and safety challenges.