Automated vehicles’ neural networks suffer from overfit, poorgeneralizability, and untrained edge cases due to limited data availability . We train two agents, aprotagonist and an adversary, using dueling double deep Q networks (DDDQNs) The coupled networks alternatelyseek-to-collide and to avoid collisions . The trained protagonist becomes moreresilient to environmental uncertainty and less prone to corner case failures than the agent trained without an adversary . The “defensive” avoidancealgorithm increases the mean-time-time to-failure and distance traveled undernon-hostile operating conditions under non-Hostile conditions .

Author(s) : Piyush Gupta, Demetris Coleman, Joshua E. Siegel

Links : PDF - Abstract

Code :

Keywords : networks - trained - time - conditions - hostile -

Leave a Reply

Your email address will not be published. Required fields are marked *