摘要 : Today, the field of practical estimation is dominated by linear-Bayesian estimators as these familiar techniques are computationally efficient and can be quickly tailored to many systems of interest. However, beyond the common wea... 展开 Today, the field of practical estimation is dominated by linear-Bayesian estimators as these familiar techniques are computationally efficient and can be quickly tailored to many systems of interest. However, beyond the common weaknesses of linear estimators to non-Gaussian noises and nonlinear transformations, Bayes' rule itself lacks robustness, as it inherently assumes that all statistical models are exactly known. Therefore, in lieu of filters outfitted with linear-Bayesian updates, this work proposes a method to derive nonlinear, non-Bayesian updates from first principles using the theory of generalized variational inference (GVI)—an optimization approach to information fusion. By selecting different divergence measures, loss functions, and feasible distributions, a wide variety of robust/conservative filter updates are quickly built and tested using numerical optimization. From these GVI candidates, a suitable update that uses the robust $\gamma$-loss function is selected based upon its desirable estimation behavior. Via analytical optimization and Gaussian mixture modeling, the $\gamma$-loss GVI update is realized as a nonlinear, closed-form filter that is well-suited for practical estimation. 收起