OgreMkV
Posts: 3668
Joined: Oct. 2009
|
| Quote (Wesley R. Elsberry @ May 27 2013,09:50) | Some time back, one of our resident antievolutionists pointed out a paper on evolvability. The research involved using evolutionary computation to assess how evolvability might change under genetic drift alone. It's a pretty nifty paper.
It also inspired me to replicate and extend the first genetic drift model they used. All the Lehman and Stanley models are done on a 2D grid. I made my program generic for dimension and I'm currently running for dimensions 2 through 5. I did this in Python, and Python 2.7.3 on my 2007-era workstation takes about an hour to do one run with the same parameters as in Lehman and Stanley. I just separated out my code to get a pure-Python GA and a separate program with analysis that uses NumPy, Pandas, and matplotlib. So I can apply PyPy to the GA part. That gets me to 19 minutes per run on the workstation.
From the first ten runs, 2D through 5D:
Just last week, Diane and I bought our first new desktop computer since 2007. I got a bundle from Newegg that gives us a 4.0 GHz eight-core AMD processor and 8 GB of RAM. The bundle included a 500GB hard drive. I added three 3TB drives to that. The 0.5TB drive I'm using for system stuff. I'm running Ubuntu 12.04 desktop on it. I added VirtualBox and set up a Windows XP VM with 512MB of memory and set aside 80GB for it. Our current file server is about out of space; it has 4TB in it. So I'm hopeful the 9TB will give us room to grow for a while. I'm planning to get the files reorganized on the new server, then use the old disks for backup. The first rsync job is underway; copying 2TB across the network looks like it will take two days.
Getting back to the evolutionary computation, I just launched five processes on the new box to handle runs of the genetic drift model. The first indication of time per run shows about 13.4 minutes. If my CPU doesn't melt, I should have the whole set of 50 runs per dimension done by tomorrow morning. I have a post up here.
I'm also looking for my stash of piezo disks so I can fabricate a couple of hydrophones. I finally figured out getting real-time clock for the Raspberry Pis and the firmware has improved to the point where I can reboot with a USB sound device plugged in, so I have the basis for a deployable audio monitoring system. There's a monitoring station I can add a device to situated on a dock in Sarasota Bay, where there's 12V DC, 120V AC, and full Ethernet access. I should be able to set up periodic recording and data upload for that station with no great difficulty. I do still need to nail down the acoustic hardware, which I intend to use a homebuilt hydrophone and amplifier kit for. What I am looking for is sufficient quality to extract numbers of snapping shrimp clicks; that should be a pretty easy target. The station samples its instruments every quarter hour, so I'm looking to do the same so I can provide results in sync with other instruments. I figure one minute of sound per quarter hour should be a reasonable first try. Once I have some results, I can see whether it is feasible to do any of the acoustic processing on board the Raspberry Pi, or whether I should handle that back home. |
How much processing power do you need and is it distributable?
I read about a grad student who built his own Beowulf cluster from Raspberry Pi boards. http://coen.boisestate.edu/ece/files/2013/05/Creating.a.Raspberry.Pi-Based.Beowulf.Cluster_v2.pdf
--------------
Ignored by those who can't provide evidence for their claims.
http://skepticink.com/smilodo....retreat
|
