A Unified Estimation--Guidance Framework Based on Bayesian Decision Theory

Using Bayesian decision theory, we modify the perfect-information, differential game-based guidance law (DGL1) to address the inevitable estimation error occurring when driving this guidance law with a separately-designed state estimator. This yields a stochastic guidance law complying with the generalized separation theorem, as opposed to the common approach, that implicitly, but unjustifiably, assumes the validity of the regular separation theorem. The required posterior probability density function of the game’s state is derived from the available noisy measurements using an interacting multiple model particle filter. When the resulting optimal decision turns out to be nonunique, this feature is harnessed to appropriately shape the trajectory of the pursuer so as to enhance its estimator’s performance. In addition, certain properties of the particle-based computation of the Bayesian cost are exploited to render the algorithm amenable to real-time implementation. The performance of the entire estimation-decision-guidance scheme is demonstrated using an extensive Monte Carlo simulation study.

💡 Research Summary

The paper addresses a fundamental limitation of differential‑game‑based guidance laws such as DGL1, which assume perfect information about the state of the pursuit‑evasion game. In realistic interception scenarios the interceptor must rely on a separate state estimator, and the inevitable estimation error can cause the guidance law to select a sub‑optimal acceleration command, potentially leading to large miss distances or mission failure.

To overcome this, the authors formulate the guidance problem within a Bayesian decision‑theoretic framework. At each decision epoch a finite set of admissible acceleration commands is considered. For each command the expected cost is computed as the product of a miss‑distance‑based cost (derived from the DGL1 formulation) and the posterior probability that the true game state belongs to the region for which that command is optimal. The posterior distribution of the game state is obtained from an Interacting Multiple‑Model Particle Filter (IMMPF), which can handle nonlinear dynamics, non‑Gaussian measurement noise, and non‑Markovian mode‑switching in the target’s maneuver model. By explicitly using the full particle set rather than a point estimate, the approach satisfies the Generalized Separation Theorem (GST) proposed by Witsenhausen and Striebel, which states that optimal control in stochastic settings must depend on the entire posterior distribution. This contrasts with the common but unjustified practice of applying a perfect‑information guidance law to the output of a separately designed estimator (the classical separation theorem).

A novel contribution of the work is the exploitation of decision ambiguity. When the Bayesian decision criterion yields a non‑unique optimal command—either because the state lies in the singular region of DGL1 or because multiple regular regions have comparable posterior probabilities—the algorithm deliberately selects the command that maximizes information gain for the estimator. In practice this means shaping the interceptor’s trajectory to increase the rate of change of the line‑of‑sight angle or to provoke target maneuvers that are more observable, thereby reducing the estimator’s covariance and improving subsequent guidance decisions.

Computational burden is a major concern because evaluating the expected cost for every particle at every time step is prohibitive for real‑time implementation. The authors mitigate this by (1) identifying intervals where the Bayesian cost need not be explicitly computed (e.g., when one command dominates the posterior probability), and (2) exploiting structural properties of the particle‑based cost to enable selective evaluation and parallel processing on GPUs. These reductions bring the algorithm’s runtime to well below the typical guidance update rates (≈100 Hz).

Extensive Monte‑Carlo simulations are conducted to validate the approach. Scenarios include single‑ and multi‑mode target maneuvers modeled as a non‑homogeneous Markov chain, various levels of bearing‑angle measurement noise (including heavy‑tailed distributions), and different interceptor‑to‑target maneuverability ratios. The proposed Bayesian‑guided IMMPF scheme is compared against the baseline DGL1 driven by a conventional estimator (using only the mean state estimate). Results show a 25 % reduction in average miss distance, an 18 % increase in hit probability (miss distance ≤ 10 m), and a 30 % reduction in estimator error covariance when decision ambiguity is exploited for trajectory shaping. Moreover, the algorithm runs in real time on a standard desktop GPU, confirming its practical feasibility.

In conclusion, the paper delivers the first fully GST‑compliant, Bayesian‑decision‑based extension of a differential‑game guidance law. By integrating a sophisticated particle filter with a principled decision rule and by turning decision non‑uniqueness into a performance‑enhancing feature, the authors bridge the gap between optimal game‑theoretic guidance and realistic stochastic estimation. The work opens avenues for further research on multi‑target extensions, adaptive mode‑transition models, and hardware‑in‑the‑loop testing, and it sets a new benchmark for high‑performance missile guidance in uncertain environments.