eduzhai > Applied Sciences > Engineering >

ShieldNN A Provably Safe NN Filter for Unsafe NN Controllers

  • Save

... pages left unread,continue reading

Document pages: 20 pages

Abstract: In this paper, we consider the problem of creating a safe-by-design RectifiedLinear Unit (ReLU) Neural Network (NN), which, when composed with an arbitrarycontrol NN, makes the composition provably safe. In particular, we propose analgorithm to synthesize such NN filters that safely correct control inputsgenerated for the continuous-time Kinematic Bicycle Model (KBM). ShieldNNcontains two main novel contributions: first, it is based on a novel BarrierFunction (BF) for the KBM model; and second, it is itself a provably soundalgorithm that leverages this BF to a design a safety filter NN with safetyguarantees. Moreover, since the KBM is known to well approximate the dynamicsof four-wheeled vehicles, we show the efficacy of ShieldNN filters in CARLAsimulations of four-wheeled vehicles. In particular, we examined the effect ofShieldNN filters on Deep Reinforcement Learning trained controllers in thepresence of individual pedestrian obstacles. The safety properties of ShieldNNwere borne out in our experiments: the ShieldNN filter reduced the number ofobstacle collisions by 99.4 -100 . Furthermore, we also studied the effect ofincorporating ShieldNN during training: for a constant number of episodes, 28 less reward was observed when ShieldNN wasn t used during training. Thissuggests that ShieldNN has the further property of improving sample efficiencyduring RL training.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×