True but I think you already know I did HMM using mixtures of Gaussians in 1986 in possibly the best set of technical papers I ever wrote at our common place of business. In a more refined version of it later, applied to Markov chains and not to continuous time and space stochastic processes, it was Desjardin's vote for paper of the year because he loudly proclaimed it so in a meeting with management. Between the several of us here with this experience, we definitely know enough to tackle this problem. That said, we are definitely going to have make this "user tunable". On my new supercomputer, with a Intel QX6880 (Core 2 Quad Extreme), and a high end GPU graphics card with its 128 little floating point supercomputers and Intel's top of the line motherboard, I will be able run at least my Markov chain on a mixture of a few gaussians since the solution for the separated Ricatti equations for the covariance is a simple 2nd order ODE and the conditional mean can be quantized in time, discretized in value to the point it applies. But if I were running a 3 GHz Sempron, I probably would give up 10 dB of perfection to get 10 dB of suppression and still run!
The "extended Kalman Filter and smoother" is particularly well suited to handling the mixture of Gaussian stochastic processes in lots of cases. When I successfully demodulated the Soviet VEGA balloon probe transmission for Ed Posner at JPL this was exactly what I used. It definitely paid to have a Cray supercomputer in those days! Now, my new supercomputer on my desk will outperform it by a lot in scalar arithmetic but with SSE4 and EMT64, in all but a few examples, it will nearly equal it in vector performance with the exception of scatter gather to the vectors of course. For shorter vectors, less than a vector of vectors, the Core 2 Quad extreme will all outperform the Cray 2 because of the huge startup latency in the memory bandwidth on the Cray. What was it Seymour said? Something like "It is a computer design I dashed off and made from personal computer parts". The memory bus frequency on 64 bit wide memory is 800 MHz on my desktop! I decided to forego the high speed memory bus technology, that which will natively support the full 1.33 GHz FSB, because Intel does not yet believe in it enough to put out a motherboard. The one I chose, takes the Core 2 Quad 1.33 GHz FSB and interfaces it to the 800 MHz memory bus seamlessly. I can live with 1.25 nsec clocks to memory! With PCI-E, I will have better I/O than we did on the Cray 2! FORGET THE GOOD OLD DAYS. We are in them. Bob N4HY Frank Brickle wrote: > On 7/11/07, *Jim Lux* <[EMAIL PROTECTED] > <mailto:[EMAIL PROTECTED]>> wrote: > > > AKA an adaptive signal canceller. You adapt to find the strongest > signal (in time/frequency, as with radar STAP), subtract it from the > input. Then find the next strongest signal, subtract it, etc. > > > That's one way, but it's not what Bob was talking about. There are > different approaches that really do rely on estimating inverse > filters. The chief differences depend on whether you can make Gaussian > assumptions about the interferers. > > They are all power independent, however. > > 73 > Frank > AB2KT > > > -- Robert W. McGwier, Ph.D. Center for Communications Research 805 Bunn Drive Princeton, NJ 08540 (609)-924-4600 (sig required by employer) _______________________________________________ FlexRadio mailing list FlexRadio@flex-radio.biz http://mail.flex-radio.biz/mailman/listinfo/flexradio_flex-radio.biz Archive Link: http://www.mail-archive.com/flexradio%40flex-radio.biz/ FlexRadio Knowledge Base: http://kb.flex-radio.com/ FlexRadio Homepage: http://www.flex-radio.com/
