Automated Testing Environment

I may have to talk to Bernhard about this, as I’m not sure what options I have, but something that could increase productivity with my PhD agent is to create an automated testing environment. This environment will consist of several domains in which the agent is tested performance-wise so I can get an idea of how well it’s doing.

The automated part comes in by this test running every week or so, without user input. All it requires is a valid agent within a CVS/SVN repository to get started. The test will also make notes on the performance of the agent for every week, thus giving feedback on how well the agent is doing.

As a side note: I am writing terribly today. I couldn’t even write this sentence properly without redoing it. And the worst part is, I need to be able to type today as I’m working on my Top Achievers Scholarship application. Fuck, it’s frustrating!

One Reply to “Automated Testing Environment”

  1. This isn’t particularly related, but I need to note down this idea.

    I’ve talked of it before, but I would like to create a human body physics simulator for an agent to learn to control. It doesn’t even have to be a human body, it could be a dog or even a 3-jointed lifeform thingy. Whatever the case, it could be something for an RL agent to learn how to control and achieve goals with. Goals could include standing up (being as high as possible without falling), jumping (this would involve correct physics), getting to a goal in 3d space (the agent would probably start by dragging itself, then perhaps crawling -> walking -> running -> sprinting).

    All this could be modeled within the RL-Glue interface, although the 3d bit may be a problem. We’ll see…

Comments are closed.