On Jan 3, 2008, at 9:03 AM, Matthew Toseland wrote:

> Useful class, but I'm really not convinced it's statistically sound.  
> You need
> to normalize the stored value after every operation, otherwise as it  
> gets
> larger, you will get bigger jumps when values are reported. In order  
> to do
> this, you probably need to extend BDRA; put some hooks in.

Hmmm... yes, in review you are right, as averaging is not linear as  
such:
assume: avg(a, b, c... x, y) = z
then: avg(a+1, b+1, c+1... x+1, y+1) != z+1

An implementation using normalized stored values is possible by adding  
a single package-hook: setCurrentValue(double d);
Then, in the averager, on report() to call  
avg.setCurrentValue(Location.normalize(avg.currentValue());

Implemented in r16855.

--
Robert Hailey


> On Wednesday 26 December 2007 20:51, robert at freenetproject.org wrote:
>> Author: robert
>> Date: 2007-12-26 20:51:38 +0000 (Wed, 26 Dec 2007)
>> New Revision: 16817
>>
>> Added:
>>   trunk/freenet/src/freenet/support/math/DecayingKeyspaceAverage.java
>> Modified:
>>   trunk/freenet/src/freenet/node/Location.java
>>   trunk/freenet/src/freenet/node/NodeStats.java
>> Log:
>> implement keyspace-aware averager

-------------- next part --------------
An HTML attachment was scrubbed...
URL: 
<https://emu.freenetproject.org/pipermail/devl/attachments/20080103/1067c949/attachment.html>

Reply via email to