6-transfer-learning

[mjadiaz]

The idea is to use transfer learning, at the most basic level as a first step, to using the trained agents in other environments:

• [x] Implement a perfect ring or circle in the models.py file.
• [ ] Implement variations of this circle.
• [ ] Use transfer learning from the full circle to the other environments
• [ ] Analysis of the policy and the q-network.

So far transfer learning for us just means continuing the training in another environment, or initialising a new run by copying the weights of the trained networks, instead of random initialisation.

Some references:

• Deep Deterministic Policy Gradients with Transfer Learning Framework in StarCraft Micromanagement

[mjadiaz]

The Circle is already implemented. Is the gaussian_2d. I think it should work as it is. We need to check if it works with the are and precision metrics.