Appendix B: Resampling Algorithms

Transcription

1 407 Appendx B: Resamplng Algorthms A common problem of all partcle flters s the degeneracy of weghts, whch conssts of the unbounded ncrease of the varance of the mportance weghts ω [ ] of the partcles wth tme. The term varance of the weghts must be understood as the potental varablty of the weghts among the possble dfferent executons of the partcle flter. In order to prevent ths growth of varance, whch entals a loss of partcle dversty, one of a set of resamplng methods must be employed, as t was explaned n chapter 7. The am of resamplng s to replace an old set of N partcles by a new one, typcally wth the same populaton sze, but where partcles have been duplcated or removed accordng to ther weghts. More specfcally, the expected duplcaton count of the th partcle, denoted by N, must tend to N ω [ ]. After resamplng, all the weghts become equal to preserve the mportance samplng of the target pdf. Decdng whether to perform resamplng or not s most commonly done by montorng the Effectve Sample Sze (ESS). As mentoned n chapter 7, the ESS provdes a measure of the varance of the partcle weghts, e.g. the ESS tends to 1 when one sngle partcle carres the largest weght and the rest have neglgble weghts n comparson. In the followng we revew the most common resamplng algorthms. 1. REVIEW OF RESAMPLING ALGORITHMS Ths secton descrbes four dfferent strateges for resamplng a set of partcles whose normalzed weghts are gven by ω [ ], for = 1,..., N. All the methods wll be explaned usng a vsual analogy wth a wheel whose permeter s assgned to the dfferent partcles n such a way that the length of the permeter assocated to each partcle s proportonal to ts weght. Therefore, pckng a random drecton n ths wheel mples choosng a partcle wth a probablty proportonal to ts weght. For a more formal descrpton of the methods, please refer to the excellent revews n (Arumlampalam, Maskell, Gordon, & Clapp, 2002; Douc, Capp, & Moulnes, 2005). The four methods descrbed here have O ( N ) mplementatons, that s, ther executon tmes can be made to be lnear wth the number of partcles (Carpenter, Clfford, & Fearnhead, 1999; Arumlampalam, Maskell, Gordon, & Clapp, 2002). Multnomal resamplng: It s the most straghtforward resamplng method, where N ndependent random numbers are generated to pck a partcle from the old set. In the wheel analogy, llustrated n Fgure 1, ths method conssts of pckng N ndependent random drectons from the center of the wheel and takng the ponted partcle. Ths method s named after the fact that the probablty mass functon for the duplcaton counts N s a multnomal dstrbuton wth the weghts as parameters. A naïve mplementaton would have a tme complexty of O ( N log N ), but applyng the method of smulatng order statstcs (Carpenter, Clfford, & Fearnhead, 1999), t can be mplemented n O ( N ).

2 408 Fgure 1. The multnomal resamplng algorthm Fgure 2. The resdual resamplng algorthm. The shaded areas represent the nteger parts of ω [ ] 1 N ω [ ]. ( ). The resdual parts of the weghts, subtractng these areas, are taken as the modfed weghts

3 409 Fgure 3. The stratfed resamplng algorthm. The entre crcumference s dvded nto N equal parts, represented as the N crcular sectors of 1 N permeter lengths each. Fgure 4. The systematc resamplng algorthm

4 410 Fgure 5. A smple benchmark to measure the loss of hypothess dversty wth tme n an RBPF for the four dfferent resamplng technques dscussed n ths appendx. The multnomal method clearly emerges as the worst choce. Resdual resamplng: Ths method comprses two stages, as can be seen n Fgure 1. Frstly, partcles are resampled determnstcally by pckng N = N ω [ ] copes of the th partcle where x stands for the floor of x, the largest nteger above or equal to x. Then, multnomal samplng s performed wth the resdual weghts: ω = ω N [ ] [ ] / N (see Fgure 1-4). Stratfed resamplng: In ths method, the wheel representng the old set of partcles s dvded nto N equally-szed segments, as represented n Fgure 3. Then, N numbers are ndependently generated from a unform dstrbuton lke n multnomal samplng, but nstead of mappng each draw to the entre crcumference, they are mapped wthn ts correspondng partton out of the N ones. Systematc resamplng: Also called unversal samplng, ths popular technque draws only one random number,.e., one drecton n the wheel, wth the others N 1 drectons beng fxed at 1 N ncrements from that randomly pcked drecton. 2. COMPARISON OF THE DIFFERENT METHODS In the context of Rao-Blackwellzed Partcle Flters (RBPF), where each partcle carres a hypothess of the complete hstory of the system state evoluton, resamplng becomes a crucal operaton that reduces the dversty of the PF estmate for past states. We saw the applcaton of those flters to SLAM n chapter 9.

5 411 In order to evaluate the mpact of the resamplng strategy on ths loss, the four dfferent resamplng methods dscussed above have been evaluated n a benchmark that measures the dversty of dfferent states remanng after t tme steps, assumng all the states were ntally dfferent. The results, dsplayed n Fgure 5, agree wth the theoretcal conclusons n Douc, Capp, and Moulnes (2005), statng that multnomal resamplng s the worst of the four methods n terms of varance of the sample weghts. Therefore, due to ts smple mplementaton and good results, the systematc method s recommended when usng a statc number of partcles n all the teratons. If a dynamc number of samples s desred, thngs get more nvolved and t s recommended to swtch to a specfc partcle flter algorthm whch smultaneously takes nto account ths partcularty whle also amng at optmal samplng (Blanco, González, & Fernandez-Madrgal, 2010). REFERENCES Arumlampalam, M. S., Maskell, S., Gordon, N., & Clapp, T. (2002). A tutoral on partcle flters for onlne nonlnear/non-gaussan Bayesan trackng. IEEE Transactons on Sgnal Processng, 50(2), do: / Blanco, J. L., González, J., & Fernández-Madrgal, J. A. (2010). Optmal flterng for non-parametrc observaton models: Applcatons to localzaton and SLAM. The Internatonal Journal of Robotcs Research, 29(14), do: / Carpenter, J., Clfford, P., & Fearnhead, P. (1999). Improved partcle flter for nonlnear problems. IEEE Proceedngs on Radar. Sonar and Navgaton, 146(1), 2 7. do: /p-rsn: Douc, R., Capp, O., & Moulnes, E. (2005). Comparson of resamplng schemes for partcle flterng. In Proceedngs of the 4th Internatonal Symposum on Image and Sgnal Processng and Analyss, (pp ). IEEE.