Hands-On Reinforcement Learning for Games

Training a meta learner

The learn2learn framework provides the MAML framework for building the learner model we can use to learn to learn; however, it is not automatic and does require a bit of setup and thought regarding how loss is computed for your particular set of tasks. We have already seen where we compute loss—now we will look closer at how loss is computed across tasks. Reopen Chapter_14_learn.py and go through the following exercise:

1. Scroll back down to the innermost training loop within the main function.
2. The inner loop here is called a fast adaptive training loop, since we are showing our network a few or mini batches or shots of data for training. Computing the loss of the network is done using the compute_loss function, as ...