Hacker News Reader

You don't need to update any weights, the model is already trained.

Ignorant here. How do you take an already trained model and execute it elsewhere? Isn’t the training phase part of the whole (ongoing) simulation?

The training phase is when you are trying to minimise your loss function by trying different weights each iteration.

Once that is finished, you have a trained model which you can use by providing input and getting an output (with the weights frozen).

The training phase is very expensive computationally because you have to calculate the gradient of your loss function on potentially huge tensors.

The execution phase is not that expensive and commercial laptops will be likely able to run the model without any problems.

My initial read of https://blog.openai.com/openai-five/ was that the 1024-cell LSTM game state layer was trained on the fly, but after rereading it may not be the case.

Reply on news.ycombinator.com