ESTIMATION-BASED SOLUTIONS TO INCOMPLETE INFORMATION PURSUIT-EVASION GAMES

Abstract

Differential games are a useful tool both for modeling conflict between autonomous systems and for synthesizing robust control solutions. The traditional study of games has assumed decision agents possess complete information about one another’s strategies and numerical weights. This dissertation relaxes this assumption. Instead, uncertainty in the opponent’s strategy is treated as a symptom of the inevitable gap between modeling assumptions and applications. By combining nonlinear estimation approaches with problem domain knowledge, procedures are developed for acting under uncertainty using established methods that are suitable for applications on embedded systems. The dissertation begins by using nonlinear estimation to account for parametric uncertainty in an opponent’s strategy. A solution is proposed for engagements in which both players use this approach simultaneously. This method is demonstrated on a numerical example of an orbital pursuit-evasion game, and the findings motivate additional developments. First, the solutions of the governing Riccati differential equations are approximated, using automatic differentiation to obtain high-degree Taylor series approximations. Second, constrained estimation is introduced to prevent estimator failures in near-singular engagements. Numerical conditions for nonsingularity are approximated using Chebyshev polynomial basis functions, and applied as constraints to a state estimate. Third and finally, multiple model estimation is suggested as a practical solution for time-critical engagements in which the form of the opponent’s strategy is uncertain. Deceptive opponent strategies are identified as a candidate approach to use against an adaptive player, and a procedure for designing such strategies is proposed. The new developments are demonstrated in a missile interception pursuit-evasion game in which the evader selects from a set of candidate strategies with unknown weights.