Confused about the underlying mechanism of Universe in measuring and training


As it mentioned :
“Universe is a software platform for measuring and training an AI’s general intelligence”.

I tried the demos in Universe and Univese-started-agent, yes, the AI can play the games successfully.

But as a new comer into the deep learning area, I am so confused what’s the underlying mechanism that Universe can help in doing “measuring and training” .

For example, as for the model in “classify traffic signs” which is a common topic in self-driving car these days, how can I utilize Universe in raining and measuring?


So here’s my unqualified answer:

Universe was built on top of the former Gym framework, and was an improvement in many ways and a lot of features where added on the back end. But it’s still a framework for training agents in model (mostly game) environments. The real magic of universe is hard to see by just training the base pong/racer agent.

It’s built modularly, in essence you have the capability to port other environments in. Then train your learning agent on whatever is determined as the reward. And the back-end can support speedy high quality data transfer. This framework is still being developed, and has already made great improvements.

check these out for more info: