Return-path: Received: from mfe1.polimi.it ([131.175.12.23]:36530 "EHLO polimi.it" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1755901AbXK0XeJ (ORCPT ); Tue, 27 Nov 2007 18:34:09 -0500 Date: Wed, 28 Nov 2007 00:29:02 +0100 From: Stefano Brivio To: Mattias Nissler Cc: "John W. Linville" , linux-wireless , Johannes Berg , Michael Wu Subject: Re: [RFC/T][PATCH][V3] mac80211: Exponential moving average estimate for rc80211_simple Message-ID: <20071128002902.5dad5804@morte> (sfid-20071127_233413_131457_C2C39227) In-Reply-To: <1196199484.8298.23.camel@localhost> References: <1196112605.8318.6.camel@localhost> <20071127163520.028f91fb@morte> <1196199484.8298.23.camel@localhost> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Sender: linux-wireless-owner@vger.kernel.org List-ID: On Tue, 27 Nov 2007 22:38:04 +0100 Mattias Nissler wrote: > Is it sensible to preprocess (i.e. calculate a smoothed average) the > input data for a PID controller? If so, should we compute a time average > or just average over the last N frames? By computing an average, you won't be able to implement differentiation, even if you could implement integration. In fact, the integration here means considering error values occurred over some time (and I'd think of some 'windowing' approach, as explained below), and so if we make an average, that doesn't really matter because all the input numbers would weigh the same in the term to be integrated, as in an average. But in order to differentiate, you need to have some array of values for the slope calculation (if you get a value only, you can't find the slope of the line which is passing by - or near to, if we consider more than two values - two points). Maybe it doesn't even make sense to have more than two values, and it would be better to just compute a regular slope between two points, so that we would just need to keep an average value and two values. Plus, the differentiation here would just mean to make a difference between the latest two values and divide it by the time difference between the two frames, so we would get just two input values. I would take this approach for computing the ID terms: [number of bad frames] [2][3][5][6][9][7] | | | 0 S <-n-> E <-- S is where the sliding window starts, E where it ends, n is the length of the window (here it's 4 as an example, that'll be subject to tuning). - in order to integrate, we make the average of the bad frame values ranging from S to E; if that average is going to be over a reasonable threshold for bad frame values we preset, the I factor will be negative, positive otherwise; - in order to differentiate, we'll just subtract the preset threshold from the values stored in E, E-1 from the preset threshold, then subtract the resulting E value from the resulting E-1 value and divide by the time which passed between the two bad frames computation; if this value is positive, the D factor will be negative, positive otherwise; - after we did these computation, we wait for some time and make the sliding window shift right, and then repeat. What we would avoid here - by properly tuning the n value - is the so-called integral wind-up [1], which could be really bad if the quality of the link varies rapidly, because we would take too much into account the past events, which could be meaningless to some degree if considering a not recent past. > Note that the rates (or actually modulations) are not a continuous > entity, but we only have a limited set of choices. Normally, the PID > would output a value from some continuous range. What are good ways to > map the PID controller output to a rate? The quick approach would be to round it to the nearest rate. A better one could be to keep a map of (1):[rate] <-> (2):[k1*rate + k2*recent errors at this rate], so that if we do have to decide whether to switch between two rates, we could actually evaluate the device performance - mainly sensitivity - at different rates(1), and accordingly think of the real difference between two rates(2). Then we round the output to the nearest rate(2) and choose the corresponding rate(1). I understand this could look a bit obscure, if you are still interested however feel free to ask questions. :) [1] http://en.wikipedia.org/wiki/Integral_windup -- Ciao Stefano