Back from honeymoon
I went quiet for a while because first I was busy getting married, and then abroad for the honeymoon. It was, of course, lovely, but now I'm back at work and getting back into my normal routine.
While not working myself, I set a number of computers to work, and in spite of a few crashes and a dead power supply I do have some useful results from that period. I'll be looking at the data in more detail over the course of this week, but the overall message is that I have found a set of conditions under which learning agents evolve relatively reliably, but they are conditions under which one run takes 5-6 days so I'm pleased but not too pleased....
Following a hunch that my advisor had, I tried increasing the duration of each presentation of food from 20 simulation time steps to 50. This has dramatically improved results, but increasing this parameter by a factor of x also increases the run time for each generation by very nearly x, hence the problem I now face. This is still progress—I have finally demonstrated that the task I'm setting the agents is at least tractable—but for practical purposes I really do need to find ways of speeding this up.
Comments
Hi!
> but for practical purposes I really
> do need to find ways of speeding this up.
Speeding up would be, of course, a good thing. But just in case it won't work too well (evolution and learning may just not be very compressible in time?) and there's a need for more processing power, then there's a new service from Amazon, called Amazon Elastic Compute Cloud (Amazon EC2). I haven't tried it and it is not free and probably requires quite a lot of additional software development by the user to get his stuff up and running, but it would be good to be at least aware of the possibility: http://aws.amazon.com/ec2
Oh, that does look useful - thanks for pointing it out.
For continuous use it would work out more expensive (>$850) to hire one machine for a year than to build one that would give me [considering that the CPU is my only bottleneck] equivalent performance, but there's a specific situation for which I could see this being a godsend: when there's a submission deadline coming up and I need to increase the sample sizes in a paper. At such a time it would be extremely handy to be able to hire a large number of computers simultaneously for a relatively short time (less than a week) so I can run loads of experiments in parallel to each other.