Boy, is somebody underclaiming:

MCM1_ 0004824_ 4390_ 1-- 732 Valid 6/8/14 04:02:29 6/8/14 08:15:25 3.56 17.5 / 80.5
MCM1_ 0004824_ 4390_ 0-- 732 Valid 6/8/14 04:02:19 6/9/14 11:46:26 4.67 143.4 / 80.5 <Me

I don't know who that wingman is, but they have fairly quick machine. To see them only claiming 4.9 points per hour is pretty strange.
Cheers

I looked at my history and I was completing those sets in 2.5 hours. I was claiming more than 17.5 points though. I have had a few WU's that I got very low points on compared to others and the wingman took about the same amount of time to complete the WU. I just chalked it up to a burp in the point calculation. At the end of the day, probably everyone will see one of those, so it evens out.

Hi i got also some lower claims during the week-end. It seems to come back to normal.
I do not know what is the reason for that ?

Those strange runtimes are likely the result of a short experiment we are conducting. As part of the process of reducing run-time variability in the MCM project, we recently sent out sent out batches of work units with modified machine-learning parameters. We expected these changes would slightly shorten work-unit run-times, but did not anticipate so strong an effect.

First, some background: processing a single work unit involves the evaluation of thousands of biomarker signatures (combinations of biomarkers). The machine-learning core of MCM analyzes each signature, but the time required to evaluate each is unpredictable. An evaluation can be stopped early, but the cost of interrupting the learning algorithm is reduced accuracy of the result.

The latest batches of WUs have their parameters modified to activate this learning-limit feature. We hypothesize that slower-to-evaluate signatures also tend to be poor quality (and therefore less interesting) signatures. If our hypothesis is true, we could use the learning-limit to achieve more stable run-times and higher throughput per work unit, with minimal effect on the most important results.

Thank you for your patience while we work to improve the run-time consistency of MCM. And as always, thank you for your interest and the CPU cycles you contribute to the project.

Christian A. Cumbaa
Research Associate
Princess Margaret Cancer Centre, University Health Network

Thanks for the feedback.