<br><br><div class="gmail_quote">2009/11/26 Kåre Särs <span dir="ltr"><<a href="mailto:kare.sars@iki.fi">kare.sars@iki.fi</a>></span><br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;">
<br><div class="im">
> perhaps you could add other statistics n'stuff too, or experiment with time<br>
> estimation methods... :)<br>
</div>Would that increase the chance of being accepted?<br>
<font color="#888888"></font><br></blockquote><div><br>Okay, so predicting remaining battery time needs artificial intelligence - the research question is to predict how many minutes of battery life remain given the current state of the machine, and knowledge of what happened in the past. Simple linear regression (reported battery remaining divided by reported power draw) is bogus for two reasons:<br>
<br>1) Accurate models of battery discharge are inherently non-linear, depending on the power draw itself, temperature, age of the battery.<br>2) Much like the stock market, current power draw alone does not accurately predict future power draw.<br>
<br>If you want to do the prediction properly, you need to learn a predictive model based on past data for each user on each computer for each battery, rather than pretending it's linear. The model would be updated periodically (say, once a minute) to account for discrepencies between what it predicted and what was actually the case, thereby gradually improving over time. The choice of prediction machinery isn't so hard (pick a form of neural net, e.g. a radial basis function), but the difficultly lies in the feature set: how, in a few numbers (to avoid the curse of dimensionality) can you characterise the current system state? If you can get this nailed, it should be possible to predict battery life with better accuracy than can be done at the moment.<br>
<br>Andrew (PhD in AI)<br></div></div>