How to Optimize a Simulator Using Meta Learning

The methods for both single task and multitask end up being quite similar. You can consider your environment parameterized by some context vector. And then you can assume maybe that once you know those parameters, you have a reasonable policy already ready to go for that set of parameters. The sort of things that people learn in the inner loop are these engine rewards is pretty big,. All of your tasks, you could have a more general parameterization of the RL objective function in the inner loops.