Deception in Autonomous Vehicle Decision Making in an Adversarial Environment

This paper considers stochastic games under imperfect information in the discretetime/discrete-state case. More specifically, the case where one player (Red) has perfect information, and the other player (Blue) has imperfect, observation-based information is considered. One motivation for this study is to demonstrate that the solution of the imperfect information game problem for the Red player collapses down to solution of an optimal control problem if Red is allowed to have knowledge of both Blue’s control algorithm and Blue’s observation process (as well as the initial state distribution). The analysis demonstrates that under the aforesaid assumptions, any Blue algorithm which maps probability distributions into some finite control space yields the same algorithmic approach for Red to compute the optimal control (which is deceptive when appropriate). The optimal control problem for Red is solved, and numerical simulations demonstrate that this optimal control for Red can significantly outperform the optimal state feedback game Red controller when deception is useful. (When there is no mismodeling, it never underperforms state feedback.) That is, the automatically computed Red controller successfully employs deception to achieve improved results where possible. Although the result is a necessary mathematical building block, the assumptions are too strong for the conclusions, on their own, to have real-world weight. In order to determine reasonableness of the conclusions, the effects of mismodeling of the Blue algorithm by Red are studied as well. The application example is a simple UAV Command and Control problem.