A Biological Inspired Improvement Strategy for Particle Filters J. P. Zhong, Y. F. Fung Department of Electrical Engineering The Hong Kong Polytechnic University joni.zhong@polyu.edu.hk Abstract-Particle Filters (PF) is a model estimation technique based on simulation. But two problems, namely particle impoverishment and sample size dependency, frequently occur during the particle updating stage and these problems will reduce the accuracy of the estimation results. In order to avoid these problems, Ant Colony Optimization is incorporated into the generic particle filter before the updating stage. After the optimization, particle samples will move closer to their local highest posterior density function and better estimation results can be produced. I. INTRODUCTION Particle Filters (PF), widely used for solving non-linear and non-Gaussian state estimation problems [1], are based on point mass particles representing the probability densities. PF are often recognized as an alternative to the Extended Kalman Filter (EKF) [2] or the Unscented Kalman filter (UKF) [3] in state estimation problems. With sufficient number of samples, PF can approach the Bayesian optimal estimate [4], rather than the EKF or UKF. However, particle impoverishment and particle size dependency are inevitably induced due to the random particles generation and uniform re-sampling applied in generic PF [5]. Some algorithms employ different sampling strategies to minimize the impoverishment, these include Binary Search [6], Systematic Resampling [7] and Residual Resampling [8]. These algorithms achieve their targets by improving the efficiency of particles. However, in the mean while, the robustness of the filtering is lost, because the diversity of particles is reduced in a certain extent [9]. In this paper, the research about the effect of Ant Colony Optimization (ACO) in eliminating these two problems is presented. We will first introduce the particle filters mechanism in next section. The ACO assisted Particle Filter together with a brief introduction of ACO will be discussed in Section III. In Section IV, experiment conducted to study the performance of the ACO assisted PF and algorithm optimizations will be presented. The conclusions are given in Section V. II. PARTICLE FILTERS Particle filters include an algorithm to perform recursive Bayesian estimation using Monte Carlo simulation and importance sampling, in which the posterior density is approximated by the relative density of particles in a neighborhood of state space. Since this sampling stage in PF is merely a suboptimal solution, therefore a major short-coming in the generic particle filters, namely particle impoverishment, is induced. A. Particle impoverishment Particle impoverishment [5] occurs when the likelihood is so narrow that the overlapping region of likelihood and the prior distribution is quite small as depicted in Fig 1.a. As a result, particle weight of particles which is far away from the region of likelihood become relatively small. Another reason for causing the problem is that the likelihood lies in the tail of the prior distribution (Fig 1.b). If such a situation occurs repeatedly, all but one sample will have negligible weights.