Risk Sensitive Particle Filters

Abstract

Abstract We propose a new particle filter that incorporates a model of costs whengenerating particles. The approach is motivated by the observation that the costs of accidentally not tracking hypotheses might be significant insome areas of state space, and next to irrelevant in others. By incorporating a cost model into particle filtering, states that are more critical to thesystem performance are more likely to be tracked. Automatic calculation of the cost model is implemented using an MDP value function calcula-tion that estimates the value of tracking a particular state. Experiments in two mobile robot domains illustrate the appropriateness of the approach. 1 Introduction In recent years, particle filters [3, 7, 8] have found widespread application in domains withnoisy sensors, such as computer vision and robotics [2, 5]. Particle filters are powerful tools for Bayesian state estimation in non-linear systems. The key idea of particle filters isto approximate a posterior distribution over unknown state variables by a set of particles, drawn from this distribution. This paper addresses a primary deficiency of particle filters: Particle filters are insensitiveto costs that might arise from the approximate nature of the particle representation. Their only criterion for generating a particle is the posterior likelihood of a state. To illustrate thispoint, consider the example of a Space Shuttle. Failures of the engine system are extremely unlikely, even in the presence of evidence to the contrary. Should we therefore not trackthe possibility of such failures, just because they are unlikely? If failure to track such lowlikelihood events may incur high costs--such as a mission failure--these variables shouldbe tracked even when their posterior probability is low. This observation suggests that costs should be taken into consideration when generating particles in the filtering process. This paper proposes a particle filter that generates particles according to a distribution thatcombines the posterior probability with a risk function. The risk function measures the