Optimal State Estimation

Chapter 15 - The Particle Filter

CHAPTER 15

The particle filter

In view of all that we have said in the foregoing sections, the many obstacles we appear to have surmounted, what casts the pall over our victory celebration? It is the curse of dimensionality, a malediction that has plagued the scientist from earliest days.

—Richard Bellman [Bel61]

We want now to point out that modern computing machines are extremely well suited to perform the procedures described.

—Nicholas Metropolis and S. Ulam [Met49]


Particle filters had their beginnings in the 1940s with the work of Metropolis, and Norbert Wiener suggested something much like particle filtering as early as 1940 [Wie56]. But only since the 1980s has computational power been adequate for their implementation. Even now it is the computational burden of the particle filter that is its primary obstacle to more widespread use. The particle filter is a statistical, brute-force approach to estimation that often works well for problems that are difficult for the conventional Kalman filter (i.e., systems that are highly nonlinear). Particle filtering goes by many other names, including sequential importance sampling [Dou0l, Chapter 11], bootstrap filtering [Gor93], the condensation algorithm [Isa96, Mac99], interacting particle approximations [Mor98], Monte Carlo filtering [Kit96], sequential Monte Carlo (SMC) filtering [And04, Cri02], and survival of the fittest [Kan95]. A short discussion on the origins of particle filtering can be found in [Iba0l]. Reference books on the particle filter include [Dou0l, Ris04].

Particle filters had their origin in Nicolas Metropolis's work in 1949 [Met49], in which he proposed studying systems by investigating the properties of sets of particles rather than the properties of individual particles. He used the analogy of the card game of solitaire. What is the probability of success in a game of solitaire? The probability may be impossible to compute analytically (because of all of the possible permutations of play). But if a person plays several hundred games and succeeds in a certain proportion of those games, then the probability of success can be approximated on that basis:

15_Optimal_State_Estimation-1.jpg

This simple idea hearkens back to the definition of probability in Section 2.1. Given the recent invention of the electronic computer at the time, Metropolis's work was certainly ahead of its time. Now that fast, parallel computers are available, his work is beginning to see its fruition in the methods described in this chapter.

As discussed in Chapter 13, the extended Kalman filter (EKF) is the most widely applied state estimation algorithm for nonlinear systems. However, the EKF can be difficult to tune and often gives unreliable estimates if the system nonlinearities are severe. This is because the EKF relies on linearization to propagate the mean and covariance of the state. Chapter 14 discussed the unscented Kalman filter and showed how it reduces linearization errors. We saw that the UKF can provide significant improvements in estimation accuracy over the EKF. However, the UKF is still only an approximate nonlinear estimator. The EKF estimates the mean of a nonlinear system with first-order accuracy, and the UKF improves on this by providing an estimate with higher-order accuracy. However, this simply defers the inevitable divergence that will occur when the system or measurement nonlinearities become too severe.

This chapter presents the particle filter, which is a completely nonlinear state estimator. Of course, there is no free lunch [Ho02]. The price that must be paid for the high performance of the particle filter is an increased level of computational effort. There may be problems for which the improved performance of the particle filter is worth the increased computational effort. There may be other applications for which the improved performance is not worth the extra computational effort. These trade-offs are problem dependent and must be investigated on an individual basis.

The particle filter is a probability-based estimator. Therefore, in Section 15.1, we will discuss the Bayesian approach to state estimation, which will provide a foundation for the derivation of the particle filter. In Section 15.2, we will derive the particle filter. In Section 15.3, we will explore some implementation issues and methods for improving the performance of the particle filter.

UNLIMITED FREE
ACCESS
TO THE WORLD'S BEST IDEAS

SUBMIT
Already a GlobalSpec user? Log in.

This is embarrasing...

An error occurred while processing the form. Please try again in a few minutes.

Customize Your GlobalSpec Experience

Category: Engineering Consulting Services
Finish!
Privacy Policy

This is embarrasing...

An error occurred while processing the form. Please try again in a few minutes.