Saddle-Point Strategy

Author: Tianjiang Says
This article is based on Zhang Chengming (2021) Research on Guidance Strategies for Spacecraft Pursuit-Evasion Games.

Definition

In zero-sum differential games, the saddle-point strategy refers to the optimal strategy combination for both players. When both players adopt the saddle-point strategy, any unilateral change in strategy leads to a decrease in payoff (for the pursuer) or an increase (for the evader). The saddle point represents the equilibrium state of the game and is the solution to such problems.

Mathematical Formulation

Zero-Sum Game Framework

Let the pursuer's control be $\boldsymbol{U}_p$ and the evader's control be $\boldsymbol{U}_e$ , with the performance index (payoff function) being $J$ . The pursuer minimizes $J$ while the evader maximizes $J$ :

\min_{\boldsymbol{U}_p} \max_{\boldsymbol{U}_e} J = \max_{\boldsymbol{U}_e} \min_{\boldsymbol{U}_p} J = J^*

Saddle-Point Conditions

The strategy pair $(\boldsymbol{U}_p^*, \boldsymbol{U}_e^*)$ is a saddle point if and only if:

J(\boldsymbol{U}_p^*, \boldsymbol{U}_e) \leq J(\boldsymbol{U}_p^*, \boldsymbol{U}_e^*) \leq J(\boldsymbol{U}_p, \boldsymbol{U}_e^*)

holds for all admissible controls $\boldsymbol{U}_p, \boldsymbol{U}_e$ .

Saddle-Point Strategy in Spacecraft Pursuit-Evasion

Problem Formulation

Based on differential game theory, the spacecraft pursuit-evasion problem is transformed into a two-point boundary value problem. According to the minimum principle, the saddle-point control strategy satisfies:

\frac{\partial H}{\partial \boldsymbol{U}_p} = 0, \quad \frac{\partial^2 H}{\partial \boldsymbol{U}_p^2} \geq 0

\frac{\partial H}{\partial \boldsymbol{U}_e} = 0, \quad \frac{\partial^2 H}{\partial \boldsymbol{U}_e^2} \leq 0

where $H$ is the Hamiltonian.

Control Equations

For the near-circular orbit pursuit-evasion problem described by the CW equation, the saddle-point control strategy is:

\sin\alpha_i^* = \kappa \frac{\lambda_{i5}}{\sqrt{\lambda_{i4}^2 + \lambda_{i5}^2}}

\cos\alpha_i^* = \kappa \frac{\lambda_{i4}}{\sqrt{\lambda_{i4}^2 + \lambda_{i5}^2}}

where $\kappa = -1$ (pursuer) or $+1$ (evader), and $\lambda_i$ are the costate variables.

Key Properties

Consistency: The pursuer and evader have the same saddle-point thrust direction
Duality: The costate variables of both parties satisfy $\lambda_p(t) + \lambda_e(t) = 0$
Reachability set boundary: The trajectory corresponding to the saddle-point strategy lies on the reachability set boundary

Solution Challenges

Initial Value Sensitivity

The solution of the two-point boundary value problem is highly sensitive to the initial guess of costate variables. Different initial values may lead to convergence to different solutions or divergence.

Computational Efficiency

Traditional numerical optimization methods (shooting method, Newton iteration, etc.) have relatively low computational efficiency, making them difficult to satisfy real-time application requirements.

Non-Uniqueness

The same initial state may correspond to multiple feasible costate variable combinations. Normalization and other methods are needed to eliminate the non-uniqueness.

Advances in Solution Methods

Costate Variable Normalization

By normalizing the costate variables, infinitely many solution sets are mapped onto a unit sphere, yielding a unique representation.

Deep Learning Methods

DRD algorithm: Uses deep neural networks to fit the mapping from initial states to saddle-point solutions
DNN-pseudospectral method: Neural network outputs serve as initial guesses for pseudospectral methods, accelerating convergence

Combined Optimization Methods

Combining traditional optimization algorithms (genetic algorithms, sequential quadratic programming) with neural networks to improve solution accuracy and efficiency.

Relationship with Game Theory Concepts

Concept	Meaning
Saddle point	The equilibrium solution of the game
Minimax principle	Evader maximizes, pursuer minimizes
Payoff function	Function of pursuit-evasion time
Strategy	Pursuit/evolution control sequence

Application Value

The saddle-point strategy is the core solution concept for spacecraft pursuit-evasion games, providing a theoretical foundation for:

Optimal maneuver strategy design in space confrontation
Approach trajectory planning for non-cooperative targets in rendezvous and docking
Guidance law design for missile interception

References

Isaacs R. Differential Games: A Mathematical Theory with Applications to Warfare and Pursuit, Optimization and Control[M]. John Wiley & Sons, 1965.
Zhang Chengming. Research on Guidance Strategies for Spacecraft Pursuit-Evasion Games[D]. National University of Defense Technology, 2021. [in Chinese]
Basar T, Olsder G J. Dynamic Noncooperative Game Theory[M]. Academic Press, 1999.